What you need

• Tesseract 0.11.37 or later. The app updates itself.

1. Ask the way you’d ask a friend

Type the question into any conversation. Some examples from the release:

• “Three-bedroom houses for sale in Austin”
• “Remote developer jobs in Berlin”
• “The best-rated sushi near me”
• “Where do these headphones cost least?”
• “What do people complain about in this app’s reviews?”
• “Is interest in this growing?“

2. Know where it looks

Depending on the question, the agent searches Google Flights and Hotels, Google Shopping, Amazon, eBay, Zillow, Google Jobs, the app stores, Google Lens, Yelp, Tripadvisor and more. 0.11.37 added ten more: Airbnb, Target, Google’s product pages, a single event in full with every ticket seller, eBay sellers, YouTube channels, Duck.ai, Naver’s AI answers, Yahoo’s trending topics, and whole websites. Concerts and local events and Bing’s product pages work again.

3. Start from a photo

Got a picture of a lamp you like? Ask what’s in the photo and where to buy it. A photo search shows the matches first.

4. Check a competitor

Ask to see a competitor’s Google ads, or what people say about their app in its reviews. It’s the same question you’d ask a colleague, answered from the sources.

What it doesn’t do

It searches and compares. It doesn’t book, buy or apply for you.

Related

When a site needs you signed in, see Let the agent sign in for you, without seeing your password. Tesseract is free to download.

What the rule does

• Only exact paths. A delete has to name the real, full path of something that exists.
• No blind wildcards. A command that would delete whatever happens to be in the folder it runs in, such as rm -f * after a cd, is refused.
• In every mode. Chat, Plan and Full access alike. Full access doesn’t switch this off.
• No interruption for you. The agent is told why the command was refused, rewrites it with the real path and carries on, without asking you anything.

What Full access still asks

Each conversation runs in Chat, Plan or Full access, and you choose which. Since 0.11.36, Full access means full access: with Check risky tool calls turned off, it no longer stops on Claude Code’s own safety checks to ask you. Two things still ask:

• A rule you set yourself.
• Anything in Chat or Plan.

With Check risky tool calls on, the agent asks you before a risky tool call.

What the agent never turns on

The agent can do a lot in the window, but it never gives itself more access. Turning on Full access, the internet, or keeping going while you are away is always your switch. For what it can do in the window, see Ask the agent to arrange your Tesseract window.

Related

The exact-path rule is on Mac and Linux. Tesseract is free to download.

What you need

• Tesseract on each computer you use: Mac, Windows or Linux.

1. Copy here, paste there

Copy as you always do on one computer. On the other, paste. Text and images both travel.

2. Find what you copied earlier

Press ⌘⇧V on a Mac or Alt+Shift+V on Windows and Linux to open Paste history. It keeps your last 500 items, with text up to 100,000 characters and images up to 10 MB, and you can search it by what you copied. Pick an item to paste it, or paste it as plain text.

It skips password managers and secure fields, so passwords you copy don’t land in the history.

3. Send the bigger things

For files, send a drop to your own devices: up to 2 GB per drop. You can also send to other people, with codes and rules for what you accept, and drops to other accounts are encrypted before they leave your machine.

Since 0.11.37 you can also just ask the agent: “Send this file to my desktop.” A file for someone else waits in the Send sheet for you to press Send.

4. Keep a whole folder in step

If you move the same project between two machines every day, point both at the same folder and Tesseract keeps the copies in step, in both directions. It skips the files your project’s git ignores and shows the state of every synced folder. Two limits: both machines need to be online, and the newest change wins. More on the page for professionals.

Related

Want to reach those computers by name, too? See Reach your own computers by name, from anywhere. Tesseract is free to download.

What you need

• Tesseract 0.11.37 or newer (Settings → About → Check for updates).
• The password saved in Tesseract, in one of two places:
  • the Vault, Tesseract’s password manager, or
  • Tesseract’s browser, saved with Save password or brought over from Chrome.

1. Save the login

In the Vault. The Vault keeps passwords, passkeys and 2FA codes, in the app and in your browser. Set a master password and they are encrypted on your device with Argon2id and AES-256-GCM, so Tesseract cannot read them. More in the privacy policy.

Or in Tesseract’s browser. Sign in to the site once and choose Save password, or bring your passwords over from Chrome. Before 0.11.37 the agent could only use the Vault’s; now these work too.

2. Ask for the task, not the login

Ask for what you want done on the site, as you would ask an assistant:

• “Sign in to the supplier portal and download last month’s invoice.”
• “Log in to the bank and tell me this week’s balance.”

When the page asks for a login, Tesseract fills in the saved one. It types the password into the page itself; the agent sees that a login happened, not the password.

3. What stays hidden

• The password itself. When the browser fills a saved login on a page the agent is working in, the password is hidden in everything the browser tells the agent, as it already was for signing in from your vault.
• Your paste history. Tesseract’s clipboard history skips password managers and secure fields, so a password you copy doesn’t end up there either.

Limits worth knowing

• This covers passwords Tesseract has: in the Vault or saved in Tesseract’s browser. A password that is only in another app has to be saved in one of them first.
• Want it in your own signed-in browser instead? That is what the browser companion is for.

Tesseract is free to download for Mac, Windows and Linux.

The same release lets you ask two more things you couldn’t ask before: how much a conversation has cost, and how full its memory is.

What you need

• Tesseract 0.11.37 or newer (Settings → About → Check for updates).
• Something you made in an earlier conversation: a logo, a deck, a page, a file.

1. Find what you made

Ask in plain words, the way you would ask a colleague who was there:

• “Where’s the logo I made yesterday?”
• “Open the deck about the spring menu.”
• “Find the spreadsheet comparing flour suppliers.”

The agent looks in Your work, the place Tesseract keeps what your conversations made, and opens it where you left it. You don’t have to remember the conversation’s name or scroll back through it.

2. Ask what a conversation cost

Tesseract itself is free. What costs money is the AI behind it, and you can bring your own subscription (how pricing works). So the useful question is how much a piece of work used. Ask:

• “How much has this conversation cost?”

It is a quick check before you start the next long job in the same conversation.

3. Ask how full its memory is

Long conversations fill up the model’s memory (its context window). Ask:

• “How full is your memory?”

If it is nearly full, that is a good moment to start a fresh conversation for the next task.

4. Ask what it can do

From the same release, “What can you do?” gets the whole answer. The agent answers across everything it can do: every skill it has, grouped by the kind of work (travel and shopping, decks and documents, videos, websites, spreadsheets, your mail and calendar, and the rest), your own skills included, alongside what the app itself does.

Related

These questions sit well with the other thing 0.11.37 added: the agent can also arrange your Tesseract window, open screens and start or move conversations, as your own clicks would.

What you need

• Tesseract 0.11.37 or newer (Settings → About → Check for updates).
• Atrium, Tesseract’s team chat. It is free with your Tesseract account, built into Tesseract or as its own app on Mac, Windows and Linux. More on the Atrium page.

1. Ask what you missed

Ask the way you would ask a colleague:

• “What did I miss in Atrium?”
• “Catch me up on #orders.”
• “What did Leo say about Café Rosa’s order?”

The agent reads the channels, threads and messages and answers in the chat. Nothing is marked read, so Atrium still shows you what is new when you open it yourself.

2. Ask follow-up questions

The answer is a starting point. Ask about the part that matters to you: “Was the rye order confirmed?” or “What did Sam say about Thursday?” Atrium has its own Ask and Catch me up too, if you would rather ask from inside a channel.

3. Reply without leaving the chat

Say “Tell Leo I’ll be late,” and the agent puts the reply in Atrium’s composer. It does not send it. You read it, change it if you like, and press Send. That rule holds everywhere the agent acts for you: anything that speaks for you or lets more in stays your decision.

4. Check on your assistants too

From the same conversation you can ask how your assistants are, what they did today, or what is waiting on you across your conversations, and pause one of their routines if you need to.

Where this fits

• Good for: coming back from time off, a busy channel you only follow, or a quick “anything urgent?” before a meeting.
• Not a replacement for reading: for a decision that matters, open the thread itself. The agent’s summary tells you where to look.
• Atrium on phones: iPhone and Android apps are coming, according to the Atrium page.

Want the agent to arrange the window while you catch up? See Ask the agent to arrange your Tesseract window. Tesseract is free on Mac, Windows and Linux: tesseract.dev.

To update from inside the app: Settings → About → Check for updates.

Every framework starts on older Linux

In 0.11.37, two helpers that Tesseract runs each conversation under needed a newer system library than older Linux systems have. On Debian 11 and 12 and Ubuntu 22.04 and earlier, that meant no framework could start. Both helpers are now built to run on any Linux.

Updates install while the screen is locked

A downloaded update now installs when your computer is idle with its screen locked, as it already did when you stepped away. On Linux it used to keep waiting.

Your private network changes at once

Joining a network, switching it on or off, or leaving it now takes effect in the running app, with no restart. In the terminal, tm peers shows each link’s path and how long it has gone unanswered. New to the private network? Start with Reach your own computers by name, from anywhere.

A question to you waits for you

When the agent asks you something, the question now stays in the chat for you. It is never sent off as a background job, and a message you type there answers it.

Delete a message on every framework

Deleting a message, with everything after it, now works on every framework: Kimi, Qwen, MiniMax, CodeBuddy, Grok, Cursor, Hermes, OpenClaw, NanoClaw and Antigravity, and on OpenCode 2.

Assistants online keep themselves up to date

An assistant you put online now updates itself while it is idle, without stopping. It keeps what it installed, its settings and its work across updates, and how each kind of assistant should work now comes with Tesseract and is updated with it.

Two features now start switched off

Keep going when I’m away and Check its work are still being finished, so they now start off, in Settings and in every conversation, including ones where you had turned them on. If you want either, turn it back on in Settings → Agent or in a conversation’s mode menu, and it stays on.

Smaller changes

• A conversation you reopen says, in plain words, that it is coming back.
• Links Tesseract opens in your browser carry utm_source=tesseract.dev, added as they open and never written into your content.

The full list is in the app under What’s New, and in the 0.11.38 release notes. Not on Tesseract yet? It is free, on Mac, Windows and Linux: tesseract.dev.

What you need

• Tesseract 0.11.37 or newer. To update from inside the app: Settings → About → Check for updates.
• Any conversation. Reading what is open works even in Chat.

1. Open what you need without leaving the chat

Ask for the file, the screen or the Settings page by name, the way you would tell a colleague:

• “Open the README of this project.”
• “Show me the frameworks page in Settings.”

The agent opens it in the window, exactly as if you had clicked your way there.

2. Lay out the window

Ask it to split the view, open a new window, zoom in or out, or go full screen. Splitting is handy when you want the chat on one side and the file it is talking about on the other.

3. Run your other conversations from this one

A conversation can start a new conversation, open an existing one, rename it, fork it, or close it. It can also move a conversation to another framework or model. For example:

• “Fork this conversation so I can try a different approach.”
• “Rename the conversation about the menu to Spring menu.”
• “Move the review conversation to another model.”

Tesseract runs 14 agent frameworks, Antigravity CLI the newest, so moving a conversation can mean a different agent as well as a different model.

What stays yours

• More access. Turning on Full access, the internet, or keep going when you are away is never done by the agent. You switch those on yourself.
• The conversation it is in. It never closes the conversation it is working in.

Your other machines too

The same release lets a conversation check your private network, switch it off or rename a computer. Anything that lets more in, like joining a network or sharing a screen, stays yours to do.

Try it

Open Tesseract, start a conversation and ask it to split the view and open the file you are working on.

How Tesseract runs it

Tesseract runs Antigravity CLI through the official server Google publishes for editors and other apps. Tesseract installs and updates it for you from Google’s own release, or uses the copy already on your computer.

Antigravity CLI runs with your own sign-in in Antigravity, or with a Gemini API key. Which accounts and plans Google allows for which uses is Google’s to say: its Antigravity FAQ has the details. Tesseract itself is free and adds no charge on top.

Set it up

1. Update Tesseract to 0.11.37 or later. It updates itself; to check now, open Settings → About → Check for updates.
2. Pick Antigravity CLI in Settings → Frameworks, or on the first-run screen. Tesseract installs it.
3. Start a conversation and choose Antigravity CLI in the framework menu next to the message box. The menu changes that conversation only.

Each conversation keeps its own framework, so Antigravity CLI can work in one while Claude Code and Codex work in others. You can also move a conversation from one framework to another, and it carries everything said so far.

What works with it

From Tesseract’s own framework catalog, on the Antigravity CLI page:

• It asks you before a risky tool call.
• You pick the model for each message and choose how hard it thinks.
• You can send it images and files, and it can drive a real browser.
• It uses Tesseract’s tools, your MCP servers, skills and extensions, and subagents.
• You can pick up a past conversation, fork one, or restore files to an earlier message.
• It shows a reviewable diff of everything it changed.
• It shows how full its context window is, and compacts a long conversation to keep going.
• It keeps working when you step away, runs scheduled, repeating jobs and checks its own work before it finishes.
• You get an integrated terminal, and you can search everything it has said.

Not with Antigravity CLI yet: running inside the operating system’s sandbox, and the cost of every conversation.

Around it

Everything else in the window is the same for every framework: the editor with a file tree and tabs, the terminal, a live preview, Chat, Plan or Full access for each conversation, and a browser the agent can drive. That is what we mean by an agentic IDE.

Try it

Download Tesseract, pick Antigravity CLI in Settings → Frameworks, and start a conversation with it.

What you need

• One account, to create the network. Only the person who creates it signs in.
• Mesh on each computer, either the free Mesh app or Tesseract itself, which has it built in.
• For a server with no screen, Node 20 or newer, to install the free command-line agent.

1. Create a network

Open Mesh, sign in, and create a network. You get a join token: one string that every machine uses to join. It keeps working until you replace it, so you can add a machine next month with the same token.

2. Join your other machines

On each computer, open the This machine tab, paste the token and choose Join. That machine needs no account. The first time, Mesh installs a small network helper and asks for your admin password once.

A server with no screen joins from the terminal, with the free Tesseract agent:

npm i -g tesseract-agent "--allow-scripts=tesseract-agent,ffmpeg-static,node-datachannel,@homebridge/node-pty-prebuilt-multiarch"
tesseract-agent mesh join <TOKEN>
tesseract-agent mesh list     # see the network
tesseract-agent mesh down     # take this machine off it

It keeps the network up as a background service.

3. Use them by name

Every machine is now reachable by its hostname from every other, wherever each one is: at home, at the office or in a data centre. Open studio-imac:3000 in a browser to see the site running on your iMac, or ssh cupboard-server from your laptop.

Traffic goes straight from one machine to another whenever the network allows. When it doesn’t, a relay passes it on, still encrypted, so your machines stay connected through almost any router. Our server issues each network’s certificates; the relay can’t read your traffic, but it does see how much passes and when.

Decide who is in

• Approval: require approval for new machines. A machine that joins waits until you approve it.
• Remove: see every machine in the network and remove one at any time.
• New token: replace the token, and the old one stops letting machines in.
• Several networks: a machine can be in more than one, and you can switch each on or off without leaving it.

A network holds up to 65,536 machines.

Send your internet traffic through one of your machines

When the network’s owner allows it, one machine can be an exit node: tick Route all my internet traffic through this network on another machine, and its internet traffic leaves through yours. A Windows machine can be an exit node after you turn on its Containers or Hyper-V feature and restart.

Inside Tesseract, the network does more

In Tesseract the same network lets you browse every machine’s files as if they were local, open their screens with Remote Desktop, and run AI agents on whichever computer suits the job. Since 0.11.37 you can also ask the agent in a conversation to check your private network, switch it off or rename a computer. Joining a network stays your decision.

More on running your computers as one is on the page for professionals.

Try it

Download Tesseract Mesh, create a network and join your second machine. If you already use Tesseract, Mesh is built in.

What Tesseract is

Tesseract is a free AI agent app for Mac, Windows and Linux. You describe what you want in a sentence — or say it — and it shows you a plan, uses the tools it needs on your own computer, and hands back real files: presentations as PowerPoint or PDF, documents, websites ready to publish. Everything it makes opens in an editor, so you can change it yourself.

Much of the desktop app works without an account. Signing in adds what needs one: a bigger free allowance, and Atrium, Tesseract’s own team chat — channels, direct messages and threads, on your computer and your phone.

Thirteen agents in one window

Tesseract runs the agent frameworks developers already use, side by side: Claude Code, OpenAI Codex, Gemini CLI, OpenCode, Kimi Code, Qwen Code, MiniMax Code, CodeBuddy, Grok Build, Cursor, Hermes Agent, OpenClaw and NanoClaw. You pick a different one per conversation — Claude Code for one task, Codex for the next — with a code editor, a terminal and a live preview beside the chat, and a plan before anything runs. That is what we mean by an agentic IDE; each agent has a page on how to sign in and what works with it, starting with Claude Code and Codex.

What it costs

There are no subscriptions.

• Bring your own AI. Sign in with the Claude or ChatGPT plan you already have and Tesseract runs on it without ever metering you. A Gemini API key works too.
• The free allowance. No AI subscription? AI is included through Tesseract, and the allowance is bigger once you sign in.
• Top-ups, from $10, only when you want the most capable models.

The pricing section has the details.

Also new in 0.11.36

• Files are deleted only by their exact path. On Mac and Linux, on every framework, the agent may delete a file or folder only by its full path, and only one that exists. A command that would delete whatever is in the folder it happens to run in — rm -f * after a cd — is refused in every mode, Full access included; the agent is told why, rewrites it with the real path and carries on.
• Full access means full access. With Check risky tool calls turned off, Full access no longer stops on Claude Code’s own safety checks to ask you. A rule you set yourself still asks.
• A message sent mid-turn keeps the reply. Typing into a conversation while it works steers it, without clearing its plan or dropping the rest of the reply it was writing.
• No more “Reconnecting” notices. When a framework retries a call to the model, the turn simply carries on.
• The audit log, in words. Team settings → Audit log now says what happened, and filters by the kind of change.

Get started in a minute

1. Download Tesseract for Mac, Windows or Linux — it installs and updates itself — or open the web version.
2. Choose Continue with Google, or Create an account with your email.
3. Sign in with your Claude or ChatGPT plan, or start on the free allowance.
4. Say what you want made.

Already using Tesseract? It updates on its own, or go to Settings → About → Check for updates. Questions, or something that doesn’t work the way you expect? Write to us and we answer by email.

One menu for everything

Press ⌘P (Ctrl+P on Windows and Linux), or the search button in the title bar. One menu opens, and it replaces the separate finder and command palette.

1. Press ⌘P.
2. Before you type, the menu shows the makers as pictures, your latest conversations and quick actions. Pick one to start.
3. Type part of a name to find a conversation, a file, one of your computers or a Settings page.
4. Type what you want to do to find the action that does it.
5. Or type a question and press Enter. It is asked in a new conversation.

⌘P inside the app, Spotlight over any other

⌘P works inside the Tesseract window. Spotlight is the shortcut for everywhere else: ⌥Space on Mac, or Alt+Shift+Space on Windows and Linux, opens Tesseract over any app, with a picture of your screen attached.

Use ⌘P to get around Tesseract, and Spotlight when the thing you are asking about is in another app, like the email above.

A moved conversation keeps its tool calls

You can run a different framework in each conversation, and move a conversation from one to another. That is a big part of what makes Tesseract an agentic IDE. Until now, a moved conversation told the new framework which commands ran, which files were touched and what the latest steps returned. From 0.11.35 its earlier tool calls arrive as tool calls, each with what it returned. That covers a move to Claude Code, OpenAI Codex, OpenCode, Gemini CLI, NanoClaw, Qwen Code, CodeBuddy, Grok, Kimi or Hermes, and a conversation reloaded on Codex to free context. The agent can rely on those calls as its own history, and the conversation shows them as tool cards, not as text.

1. Open the conversation you want to move.
2. Pick another framework in the framework menu next to the message box.
3. A line above the message box says where the conversation is going and how many messages it carries. Anything you type meanwhile waits, and goes there once it has arrived.
4. Carry on. The new framework has the earlier tool calls and their results as its own.

Two fixes go with it:

• Only what you wrote is in your name. A conversation opened again could show messages you never wrote as yours: notices from the project’s other sessions, the result of every tool call in a moved conversation, and the note that it carried on after a model switch or a restart. They now show only with Show geek stuff, each under a name that says what it is.
• A conversation its framework lost carries on. When Claude Code, Codex or another framework no longer had a conversation’s history (“No conversation found” on every message, or a thread that disappeared), the conversation could not go on. Now it continues from Tesseract’s own copy, in a fresh session.

Also in 0.11.35

A restart in the middle of a task keeps the task. If the app restarted while the agent was still working, or waiting on a question it asked, the “continue” you type afterwards is part of that same task on every framework. Before, it was reviewed on its own.

A step you drop reads “not needed”. Tell the agent a step of a plan like this one is no longer needed, and the review now marks it “not needed”, even if the agent ticked it off as done.

• Hand-offs written from what was done. A session can write its hand-off, or one for another session in the project, from what was recorded rather than from memory: what it was asked, its uncommitted files and commits, the commands it ran and how each ended, its plan and what it waits on. It can keep it in docs/handoffs.
• Verified on more frameworks. On the Tesseract Gateway, runs on Grok Build, MiniMax Code, CodeBuddy, Hermes, Qwen Code and OpenClaw always ended “Not verified”. Their work is now reviewed like any other.
• The model you pick is the model that answers. On the Tesseract Gateway, MiniMax Code now runs the model you choose in the menu.
• Programs a framework starts close with it. A dev server a framework started no longer keeps running for hours after the framework and the app are gone.
• On Windows, Tesseract comes back if it closes. After a crash, or an update that did not restart it, a small watcher starts the app again within a minute.
• Remote Desktop: a sharp picture and one cursor. The picture sharpens a few seconds in, and only the remote cursor, where clicks really land, is shown.

Try it

Download Tesseract for Mac, Windows or Linux at tesseract.dev, then press ⌘P (Ctrl+P on Windows and Linux) and type what you are looking for.

Call someone from a team conversation

Team conversations now have Call and Video call. A call rings the other person on every device they are signed in on, with Decline, Answer with audio and Answer with video. A call nobody picks up says “No answer”, and one you miss leaves “Missed call”. When either of you hangs up, it ends for both.

1. Open the Team tab and the conversation with the person you want to reach.
2. Press Call, or Video call to start with your camera.
3. They answer on any device they are signed in on, with audio or with video.
4. While you are muted, hold Space to talk. Press ⌘⇧M (Ctrl+Shift+M on Windows and Linux) to mute.
5. Hang up when you are done. The call ends for both of you.

The Team tab is Tesseract’s built-in team chat, Atrium, with your channels and direct messages. Atrium is also a free app of its own. Voice and video calls are still in beta while they roll out. Group calls can share a screen, and live captions work with your own OpenAI or Gemini key.

Everything you made, in one place

Home now shows your work. Every assistant, presentation, website and app you made is there, each with its own picture, how it stands (online, published, how many slides) and when you last changed it.

• The ⋯ menu on each one opens, presents, visits, renames or deletes it.
• See all lets you search and sort everything you made.
• A new conversation offers the four newest under Pick up where you left off.

So you don’t have to remember which conversation made last week’s deck. Find it on Home and present it from its menu, or start a new conversation and pick it up from there.

More to read while the model thinks

When a model is thinking and has nothing to show yet, the chat now says “Thinking” with a running clock, instead of “Waiting for the model’s reply”. OpenAI’s reasoning models are also asked for a summary of their thinking on every framework, so there is more to read while they work. Codex’s thinking summaries now read as paragraphs, instead of running into each other mid-sentence.

Two smaller changes help the agent see what you see:

• Every framework knows the chat’s browser tab. The tab in the chat’s browser chip is named to the agent on every turn and on every framework, not only in Claude Code, and that includes a tab in your own browser. The agent can also open a page in a new window of your browser.
• Transparent pictures drawn by the model. With OpenAI, a picture asked for with a transparent background is now drawn transparent by the model itself, with clean edges and no coloured fringe. The image editor’s Transparent switch no longer asks for a key colour.

Follow-ups stay part of the task

With verification on, “continue”, a correction or an answer to the agent’s question is now checked against the whole task you started, not on its own. An unfinished task also survives an app update, and plan steps that are no longer needed show as “not needed”. With verification off, a run that stops halfway down its own plan is now told to carry on in every framework, as Claude Code already was.

A long job like this one now stays one task, from your first message to the last follow-up. Its file changes read as steps of its work line too, such as “Created report.md” or “Changed app.ts”, and a file’s name opens it.

And on big projects, a window now watches only the folders you are looking at: the expanded folders, the Files browser’s folder and the files you have open. Before, it watched every project tree, which on Linux meant a watch for every file.

Try it

Download Tesseract for Mac, Windows or Linux at tesseract.dev. Then open a team conversation and press Call.

Share a computer with your team

Open a computer’s page from My Computers. Under Sharing, share it with an organization you are in, and its members can use it as you do. Shared computers show whose they are, and you can make one personal again at any time.

You manage who is in an organization in Settings → Team, which now works for every organization you are in. Owners and admins change roles, remove people and invite by email. Members see who is in it and can leave.

If you’d rather a computer weren’t reachable at all, turn off Make this machine available. That switch now only stops sharing this computer. Before, it could close the app a second after opening, and leave a signed-in app unable to see your other computers.

Add a computer, or hand one over

Adding a computer now takes one code, and giving one away takes one email.

To add a computer:

1. Open My Computers and choose Add a computer.
2. You get a one-time code, shown as a QR code, a link and ready-to-paste commands for a server.
3. Use it on the new computer. The code works once, for 15 minutes.

To give a computer to someone else:

1. Open the computer’s page and choose Transfer to someone.
2. Enter their email.
3. They accept at the top of My Computers.
4. A command-line install moves to them by itself. A computer running the app moves when they sign in on it.

You can also invite a colleague by email, and change the email you sign in with in Settings → Account.

My Computers, redrawn, and one computer per conversation

My Computers now groups your computers by kind, each on one line with its system, how long it has been online and load bars. The bars turn amber or red only when a computer is nearly full. The top counts what is offline or waiting on a transfer, and Files, Terminal and Remote desktop open straight from each row.

Each computer’s page also has a new Configuration tab. It shows the framework and model new conversations use there, the account it runs on, its defaults, skills and plugins. You can change them from any of your devices.

Until now, browsing another computer’s files or opening its terminal switched the whole window to that computer, and every conversation then ran there. Now each conversation stays on the computer it belongs to, and another computer’s files open in a window of their own.

Files still browses any of your computers from the Machines list, as here on the Studio iMac. The difference is that looking around no longer moves your work.

Remote desktop is one click from a row in My Computers, and its own tab keeps a list of your computers too. You can connect with a one-time code, ask for access by email, or use a password for a machine nobody is sitting at. Remote Desktop is also a free app of its own.

Settings, cut to what works

Settings went from 22 pages to 14. Options that worked with only one framework, or did nothing at all, are gone. If you run Claude Code, Codex and OpenCode side by side, a setting you see now means something whichever one a conversation is on.

Every page groups its options in titled cards, with one line of help each. Adding an account asks only three things: where it runs, which framework, and what that pair needs.

A few more changes in 0.11.33 that you will notice:

• A calmer chat. Answers sit in a comfortable reading column, and the work behind each one folds into a single line you can open. The files a reply changed are collected in one card, each with its own Undo.
• Conversations stay in step across your devices. Replies arrive once, in full, wherever the conversation is open.
• Claude Code costs are counted once. Each answer now shows only what it cost, so totals and spending limits read right.
• When your Claude plan runs out, the chat says so and when it resets, and offers to carry on with another framework.

Try it

Download Tesseract for Mac, Windows or Linux at tesseract.dev, then open My Computers and choose Add a computer to bring in the next one.

Start with a demo you can tap

The mobile app maker has been on Home since 0.11.30. Your app first appears as a demo you tap through on a phone: any current iPhone, iPad, Pixel or Galaxy, held sideways or folded open. It starts from one of 24 ready-made app designs, or it is drawn from scratch and checked on a phone. While it takes shape:

• Write it for… researches the app’s words, screen by screen.
• Double-click words or a photo on the phone to change them.
• Point shows the agent the part you mean.
• Try any of 14 looks, and go back in History if you liked an earlier one better.

You don’t have to start from Home, either. Ask for an app in any conversation and the agent opens the maker right there, keeping what was said.

Make it a real app, now in SwiftUI too

When the demo is right, Make it a real app builds it for real. Since 0.11.30 that has meant Flutter or React Native, with every screen checked against the demo. Since 0.11.31 it can also build the app in SwiftUI, on a Mac with Xcode. The app is built and opened on the iPhone simulator, and every screen is checked against the demo there.

The checks are stricter too. Each built screen is also looked at the way a person opening it would see it. The app sits inside the phone’s safe area, fills the phone and opens each screen as its own page.

1. On Home, choose Create a mobile app.
2. Say what the app is for. The chat asks only what it still needs, while the demo takes shape beside it.
3. Tap through the demo and change what you want: double-click words or photos, use Point, try another look.
4. Use Write it for… to have the app’s words researched, screen by screen.
5. Press Make it a real app. On a Mac with Xcode it can build the app in SwiftUI and open it on the iPhone simulator. Flutter and React Native are there as before.

Conversations in one project know about each other

A project often has more than one conversation working in it, and sometimes Claude Code running in a terminal or an editor beside them. Since 0.11.31 they can see each other, tell whose each uncommitted change is and send each other a word. So one no longer commits another’s half-done work. That works for Tesseract’s conversations on any framework, and for Claude Code outside the app.

Two lists show it. Working here at the top of Conversations shows who else is in the project, and Uncommitted changes under the file tree shows whose each change is.

0.11.32 goes a step further. Sessions in the same project now claim the files they are changing, so another session cannot edit them underneath. Each one commits or undoes only its own changes, and Claude Code running in a terminal beside them can be messaged too.

A session can run a team

A conversation can now start other full conversations for parts of its work, such as a developer, a tester and a reviewer. It follows them and brings their work together. Each of them is a full conversation, and each can be on a different framework.

The menus by the message box set each conversation’s mode, model and framework, as in any other conversation.

You review a plan like this one before the work starts. When a conversation runs a team, its plan also shows which steps wait for which, and the reports that come back from the team move them along.

Agents outside Tesseract can join in as well. Codex, Gemini CLI, Cursor or an editor’s agent can join a project’s sessions through the tesseract-crew tool server. And the same project on two of your computers is one crew, not two.

Try it

Tesseract is free for Mac, Windows and Linux at tesseract.dev. Choose Create a mobile app on Home, or open a project and start a second conversation in it.

Make one in a minute

Create a personal assistant is one of the cards on Home, and it makes an assistant from your first sentence with a single tap. On a Mac it reads Mail, Calendar and Reminders, writes you a brief each morning and a weekly report, and moves your own events when you ask.

1. Make sure you are on Tesseract 0.11.30 or later. The app updates itself.
2. On Home, choose Create a personal assistant.
3. Say what you want help with in one sentence, for example: “Help me run the bakery: supplier emails, custom cake orders, the day’s deliveries and the weekly numbers.”
4. Read the set-up beside the chat: what it will look after, what it does on its own (a morning brief, a weekly report) and what it always asks you first. Change anything by saying so, then press Make it.
5. On a Mac, macOS asks once to let the assistant read Mail, Calendar and Reminders. Say yes, and its first brief follows.

An assistant can also be a specialist, for research, sales outreach, your inbox or travel, and assistants can hand work to each other. Each one has its own page with its briefs, its routines, the drafts waiting for you and Meet a future teammate. Your assistant runs on your computer, so it works while that computer is on.

Talk to it where you already talk

When you are signed in, Talk in Atrium puts your assistant in a direct message in Atrium, Tesseract’s team chat, with nothing to set up. It can also be on Telegram, WhatsApp, Discord, Slack, Signal, Matrix, Mattermost or email.

A few rules hold on every channel:

• It hears voice notes and can answer with one.
• Before it acts for anyone else, it asks you on your own channel. No answer in 30 minutes is a no.
• Messages from strangers wait for your approval, or are summed up for you.

Give it your accounts, and teach it your tasks

Connections on an assistant’s page signs in to Gmail, Google Calendar and Contacts, Outlook and Teams, and 62 services such as Notion, Linear and Stripe. Each account gets a label (“work”, “personal”), and only the assistants you allow can use it: on its channels, and when you talk to it in the app. Until Google and Microsoft finish reviewing Tesseract’s app, their sign-in page may say it is unverified.

Some sites need signing in. The agent can have Tesseract type your saved username, password or current two-factor code from the Vault, only on that site. It learns which account was used, never the value. An assistant’s unattended runs keep their sign-ins too, and save downloads into its folder.

For a job you do by hand, use Teach it a task. It records you doing it once in a browser tab, never your passwords or codes. Then it writes the job up as a skill, tries it as a dry run that does everything but act, and makes it a routine. Watch me do this works the same way in any conversation.

Your assistant can also ask around for you. Ask it to find something out from people, say which pharmacy has a medicine in stock. It first asks you what they might want to know and shows you the plan. Only after your yes does it contact them by text, email, message or phone, and Reaching out on its page shows where each one stands. Calls are held by your own live voice model (your Google key in Settings → Voice), on a Twilio number of your own or a line of your own PBX.

It asks first, and stops when you say

What an assistant does on its own runs as routines. In the activity map, a scheduled task shows its runs: how each ended, how long it took, what it cost and what it usually costs. Test tries it as a dry run that acts on nothing. A routine can also be set off by a GitHub, Sentry, Linear or Stripe webhook, any other webhook, or new mail in a mailbox, and you can pause each trigger.

When a run nobody started reaches a step that needs your OK, the step waits for you on every device you are signed in on. You get a reminder after 10 minutes, and it is declined after 30. The first answer counts, and a card answered elsewhere says where. Always allow saves a narrow rule (the command, the file or the site), listed in Settings → Permissions & sandbox → Saved approvals.

Tesseract already works this way when an ordinary conversation keeps going while you are away. In this overnight job, it made the changes it was approved for and held back deleting 40 unused photos, because nobody was there to say yes.

To cap the cost, Settings → Agent → Spending limits sets hard limits for an assistant, or for everything, one project or a team. Limits can be per conversation, day, week or month, on every framework. You are told at four-fifths of a limit, and once it is used up nothing more is sent and the runs on it stop.

Try it

Download Tesseract for Mac, Windows or Linux at tesseract.dev, choose Create a personal assistant on Home and tell it what you want help with.

How a maker works

You answer a maker’s questions in the chat. Its conversation is a full chat with every control, so you can also ask for any change in your own words.

You don’t have to start from Home. Every maker can also be started from the command palette. And if you ask for a picture, a logo or an app in any conversation, the agent opens that maker right there, or moves from one maker to another, keeping what was said.

Some cards were merged. “Design a logo”, “Create a font” and “Create a brand kit” became Create a brand. “Write a blog post” and “Write an email” became Write something. “Compose music” and “Schedule a task” left Home. The Design Studio tab is gone too, and Build a website takes its place.

The first run now asks who uses Tesseract: a kid or teen, an everyday user, or a tech expert or developer. Home’s cards come in the order you are most likely to want them, and you can change the answer in Settings → General.

Pictures, video and a brand

Generate an image asks what the picture shows and what it is for, then draws three different takes at the exact shape of where it goes, such as an Instagram post, a story, a banner or print. Every change after that is an edit you can go back on, with Compare. Change a part lets you brush over an area and say what it should become, and Also make it for… makes the same picture for other places. When words drawn in a picture come out wrong, press Fix the words or Set them as real text.

Make a video writes the script first, then makes a picture of every shot, then every shot at once on Veo while a rough cut plays. It is put together on your computer with captions, music and the narrator you picked. Use my photos turns your own photos into moving shots, and Your own voice lets you read the lines yourself. Videos are made at the Light quality (Veo 3.1 Lite) unless you choose Standard, or Best with a Google key of your own.

Create a brand makes a logo the way logo makers do it: symbols drawn for what the business does, beside the name set in real type so it is always spelled right. Then try layouts, typefaces and colours without drawing again. Save gives you the whole kit: every logo version as SVG and PNG, an app icon and favicons, a palette with each colour’s job, the typefaces, social media covers, a business card and a one-page guide.

Three more makers sit beside them. Make social posts plans a set of posts with a caption written for each platform and matching pictures. Make a postcard designs both sides and gives you a print-ready PDF. Enhance a photo plans the few fixes your photo needs and shows a before and after.

Pictures need credit or an image key of your own, and the makers say so, with the way to add either.

Research, writing, apps and games

Do deep research shows its plan of sub-questions before anything runs, and you change it by saying so. You watch it tick off the plan and can read its sources. Stop it with That’s enough — write it up, and you get a cited report to save as PDF.

Analyze data takes a spreadsheet, a CSV or a pasted table, or a sample to try. It shows what your data holds before asking you anything. You agree a plan, watch the charts arrive, and get the answer with its key figures, what was changed in your data, how many rows were used, a results workbook and the script to run it again.

Write something handles an email, a letter, a post or a CV. It shows an outline first, then writes a draft you edit directly and change by asking, keeping your own edits. Hand it over with Open in Mail, Copy, or save it as Word or PDF.

Create a mobile app shows your app as a demo you tap through on a phone, made from one of 24 ready-made app designs or drawn from scratch. Create a game asks whether it plays right here in the browser or is a real Unity, Unreal or Godot project. You see its design before anything is built, then play it on the canvas. Create a custom project sets up a folder with the exact name you give, where you choose, with the stack, files and tools you describe. It shows the plan first, checks that it runs, and offers Open it here.

Create my Tesseract skin restyles the app itself. Describe a look or give a picture to take its colours from, and see the app in it in light and dark. Change it back undoes it, and Save to share makes one .tesseract-skin file that anyone can add in Settings → Appearance → Add a skin from a file…. There is also Create a personal assistant, which makes an assistant from your first sentence.

How to start a maker

1. Open Home.
2. Press a card, such as Generate an image. You can also start it from the command palette, or ask for it in any conversation.
3. Answer what the chat asks, in your own words.
4. Watch the work take shape beside the chat. Change it by asking, or with the maker’s own buttons.
5. Keep the result: save it, export it or hand it over, the way each maker offers.

Try it

Tesseract 0.11.30 is out now and updates itself. Download it at tesseract.dev, open Home and press the card you have been curious about. The blog has more on the website and presentation makers.

Catch up, translate and ask

Team chat is built into Tesseract: channels, direct messages, threads, mentions, reactions and search, in the same window as your conversations with the agent. Coming back to a busy channel no longer means scrolling up to where you left off.

• Catch me up appears when three or more messages are waiting. It sums up what you missed, starting with what asks something of you, and every point links to its message.
• Translate, in a message’s menu, shows that message in your language. Names and links are left untouched, which helps in a team that writes in more than one language.
• Ask, in search, answers a question about the team’s history from the messages you can read. Each sentence cites its sources, so you can check it. Search still finds the words you type; Ask answers the question, using only what you could open yourself.

Voice messages and voice rooms

Some things are quicker to say. Record a voice message of up to fifteen minutes from the mic in the message box, and listen back before you send it.

What you said is written down on your own computer, and the recording is not sent to any transcription service. Writing it down there takes a one-time download of about 250 MB.

Voice rooms show who is in and who was there last. You can join muted, or on the microphone you pick, and each room has its own text chat for links and notes.

How to use the new features

1. Sign in, and open the Team tab. Team chat needs a Tesseract account.
2. Open a channel where three or more messages are waiting, and press Catch me up. Press a point’s number to jump to its message.
3. For a message in another language, open that message’s menu and choose Translate.
4. To find something out from past messages, open search, type your question and choose Ask.
5. To send a voice message, press the mic in the message box, record, listen back, and send.
6. Press ⌘J to jump anywhere, and ⇧Esc to mark everything read.

Guests, polls, and a redesign

Team chat also got the everyday tools a team leans on.

• Guests. Invite people from outside the team into a channel by email. They see only the channels they are invited to, so someone from outside can join one conversation without seeing the rest, and guests are free.
• Profile cards. A name or a face opens a card with what that person does, their status and their local time.
• Scheduled messages. The arrow beside Send schedules a message for later.
• Later. Any message can be saved for Later, with a reminder.
• Polls. A poll can be single or multiple choice, named or anonymous. Results stay hidden until you vote.
• Forwarding. A message can be forwarded to several places at once, with a note.

The whole chat has a new look, too. It now has profile photos, day dividers, a “New” line where you left off, a toolbar on each message, the full emoji set with Frequently used first, and a message box that formats as you type. Search filters by person and date and marks the words it found. A sent message is edited in the same editor you wrote it in, and code blocks have Copy.

Some long-standing problems are gone too. Since August, mentions showed as a scrap of an ID, channel links were broken the same way, times showed as ”…” and spoilers never hid. A line break typed with Shift+Enter arrived as a space. Forum and voice channels were treated as direct messages, so they could not be renamed, moved or joined. All of this is fixed.

Try it

Tesseract 0.11.28 is out now and updates itself. Download it at tesseract.dev, sign in, and press Catch me up the next time a channel gets ahead of you.

Questions in the chat, the site beside it

You don’t need to know what a good website needs, because the maker asks. Tell it about the business in your own words, and answer what it asks next.

Then choose how it looks. You get three looks to pick from, More looks if none of them fits, or My brand, which takes the colours from your logo.

From there, Tesseract researches, plans and writes each section. Where the business needs it, it adds pages of their own, such as a menu, services with prices or a schedule, with a site menu that works on a phone. The site is built in a Design Studio tab beside the conversation, so you can watch each section arrive.

The agent’s first drafts are better too. It can have a design scored and keep revising it while each new version beats the best so far, and it can find real photographs for it on Unsplash or Pexels. Its first drafts now follow what the judges kept asking for; in blind comparisons, the new ones were preferred in 8 of 12 cases.

Websites can also start from Tesseract’s templates, and many of them got better. In all, 115 templates were fixed so their words can be read on photos, on coloured buttons and in faint greys, at desktop and phone width. Eighty-five now lay out properly on a phone, with nothing covering their words or running off the screen. The crisis-help template keeps its Quick Exit button on a phone.

Change it by pointing at it

You don’t have to describe where something is. On a website’s canvas:

• Double-click any words to change them in place, with Undo.
• Double-click a photo to swap it for one of yours, a real one found for it, or one drawn for it.
• Point puts any part of the page into your message, so the agent knows exactly which part you mean.
• Look tries colour and type styles as you hover over them.
• History goes back to any earlier version.

Anything you would rather explain in words, you can still type in the chat.

How to build one

1. On Home, press Build a website, or choose New website in the command palette.
2. Say what the business is in a sentence or two, and answer the questions in the chat.
3. Pick one of the three looks, press More looks, or choose My brand to use your logo’s colours.
4. Watch the sections and pages being written beside the chat.
5. Change what you want: double-click words or photos, use Point to show the agent a part of the page, try styles with Look, and go back with History if you change your mind.
6. When it’s right, press Make it a real website and choose Next.js, Astro, SvelteKit, Nuxt or Django. If its checks find problems, press Fix them.

From a draft to a real website

Make it a real website turns the site you shaped into a project built in Next.js, Astro, SvelteKit, Nuxt or Django: a code project of the kind web developers work with, which you can keep and build on. Its checks look the result over, and Fix them mends what they find.

Like everything Tesseract makes, the site’s files open in an editor, so you or a developer can keep working on them by hand, next to a live preview and a terminal. If you work on code, the agentic IDE page shows the rest of that workspace.

This release also fixes two things in the Design Studio. A website built on OpenCode could never reach Run checks: it waited for a plan card OpenCode never shows, because there you approve by switching to a building mode. That switch now counts. And an open Design Studio tab and its build conversation now come back after a reload or a restart.

Try it

Tesseract 0.11.28 is out now and updates itself. Download it at tesseract.dev, press Build a website, and describe a business you know.

What changed

On Codex, OpenCode, Gemini CLI and the other frameworks besides Claude Code, the agent can now:

• make images and videos
• draw charts on the canvas
• write reports
• find your earlier conversations
• ask for an API key on a masked card
• publish a page
• set a goal
• show cards side by side

Until this release, only Claude Code could. Your skills reach every framework too.

This finishes a move that started in 0.11.26. That release let one conversation run on Claude Code while another runs on Codex, at the same time, and brought forms, choice pickers and the plan checklist to Codex, OpenCode and Gemini CLI. It also let the agent on those frameworks hand work to a helper on another installed framework or model.

What each new tool does for you

Images and videos. Ask for a picture or a video in any conversation, on any framework. Pictures can also be edited with layers in Tesseract’s image editor.

Charts and reports. The agent can draw a chart on the canvas, or write a report, whichever framework the conversation runs on.

Your earlier conversations. The agent can find and read your earlier conversations, so you can point it at work you did before instead of explaining it again.

An API key, kept out of sight. When a job needs a key, the agent asks for it on a card with a masked field and uses it without ever seeing it.

A page or a file at a public link. When you ask, the agent can publish a page it built, or share a file at a public link.

A goal. The agent can set a goal whose criteria are checked before the work is called done.

Cards side by side. It can show cards or images next to each other in the chat, which helps when you are choosing between options.

How to try it on another framework

1. Open a conversation, or start a new one.
2. Use the framework menu next to the message box to pick Codex, OpenCode, Gemini CLI or another framework you have. The menu changes that conversation only, and it carries everything said so far across.
3. Ask for one of the things above: a picture, a chart, a report, or the conversation where you worked on something before.
4. If the job needs an API key, type it into the masked card the agent shows you.
5. Look at the status bar while it works. It now names the framework that is working, where it used to say “Claude is working” whatever the conversation ran on.

Also better on every framework in 0.11.27

• Presentations. Once a presentation is built, the assistant is told which file each slide is, which slide you are looking at and what each research note is about, so “add this to this slide” is done where you meant it. On OpenCode, Gemini and the others this never reached the assistant before, and on Codex only with a conversation’s first message. The plan, the slides and the read-through now use the model your conversation runs on, on every framework.
• The Companion side panel. The chat in the Companion extension’s side panel now has a Run on menu with every framework and account this computer runs, each framework’s own models, and Chat, Plan and Full access.
• OpenCode 2. On a computer with OpenCode 2 installed, every OpenCode conversation failed to start. OpenCode 2 now starts and runs your conversations, including the extra folders a conversation was given. Forking from a message and deleting a message don’t work on it yet.
• OpenCode’s steps. A step used to say only “bash” or “apply_patch”. Each step now fills in with the command it ran or the change it made, as soon as OpenCode sends the details.
• Two OpenCode conversations in one folder now keep to themselves. Before, opening the second could cut off a call the first was making.
• Two Codex conversations started at the same moment both get Tesseract’s tools. Before, the second could run without questions, plan or helpers.

Try it

Tesseract 0.11.27 is out now and updates itself. Download it at tesseract.dev, switch a conversation to the framework you like best, and ask it for a picture.

Research first, saved as notes

The Make a presentation card on Home still starts with a short conversation: what it is about, who will watch and which look you like. What happens after your answers is new.

Tesseract looks into the topic with the conversation’s own framework, whether that is Claude Code, Codex or Gemini CLI, on the web where it can. Every fact it keeps goes into the presentation’s notes with where it comes from and how far it can be trusted. The notes appear beside the chat as they are saved, and stay one click away afterwards.

You can steer the research while it runs:

• Type something, and it is looked into too.
• Say that’s enough, and it plans from what it has.
• Attach your own material, and it is read first. What you attach outranks anything found elsewhere.

A pitch, a sales deck, a proposal or an update is your own case, and it is now made that way. The research reads your own material first: the project you are working in, your earlier presentations and what you attach. The plan and the slides then speak in your voice (“we”, “our”) instead of reading like someone else’s report on you.

A plan built on evidence

The plan now starts from the one point the whole presentation makes. Every slide gets a title that says its own point, and you can see how many facts each slide rests on. What the notes could not prove is looked into once more before you see the plan. What only you know, such as your own figures, names and dates, is listed for you to add.

There is also a better question about length. Instead of a number of slides, you say how you will use it: present it for 5, 10, 20 or 30 minutes, at about a slide a minute, or send it to be read. A talk gets slides to glance at and notes to speak from. A presentation that is read gets fuller pages that stand on their own. Long talks get section slides, and a plan can now hold up to forty slides. The minutes you give are held to, also after you add slides to the plan.

Slides that show their point, and notes you can speak from

Each content slide now carries one visual that fits what it says: bars, a ranking, a trend, a share, a few key figures, a funnel, a small table, a comparison, a timeline, steps, a cycle, a pyramid, a matrix and more. There are twenty-one in all, drawn in the presentation’s own look. Numbers come from the research, are written the way your language writes them, and the slide says where they come from.

Each slide’s speaker notes are now what to say:

• the point, in one sentence
• what the audience is looking at
• something that is not on the slide
• a line that leads to the next slide
• a fact kept in reserve for the question that will come

The opening and the close are written as an opening and a close.

When the slides are written, the whole presentation is read once more for a point made twice, figures that disagree and a claim its slide does not back, and those slides are rewritten. Slides whose figures have no source are marked for you to check. The plan, the slides and this read-through all use the model your conversation runs on, not its smaller fast model.

How to make a researched presentation

1. On Home, press Make a presentation and say what it is about. Attach a Word file, a PowerPoint, a PDF or your notes if you have them.
2. Answer who will watch, how you will use it and which look you like. Tap an answer, type it in your own words, or press Decide for me.
3. Watch the research notes fill in beside the chat. Type anything else you want looked into, or say that’s enough.
4. Read the plan, and add the figures, names and dates it asks you for. Open Read and edit under the plan to see the notes: press Edit on a note to change it in place, add a note of your own with what you know, or remove one.
5. Build it. Facts under the open slide shows what that slide is written from. Take out a fact that is wrong, add one of your own, and say in your own words what else the slide should say or show. Rewrite this slide writes just that slide again and leaves the rest as it is.

Later, the notes are also under Research notes in the presentation’s menu.

When the notes change, the slides can follow. Ask in the chat for more information, such as “look this up” or “ground it in my project”, and the assistant adds it to the notes. Tesseract then tells you what changed and offers to update the presentation. It plans it again from the notes and shows you the plan first. Nothing in your slides changes until you build it, and Leave my slides as they are puts everything back. What you asked the assistant to change on a slide is kept when the presentation is built again.

A web page, PDF or PowerPoint made from a presentation carries its slides, never its research notes. In a PowerPoint file, figures that count up on screen are now written in at once, so the number in the file is always the number.

Try it

Tesseract 0.11.27 is out now, and the app updates itself. Download it at tesseract.dev, press Make a presentation and pick a topic you know well, so you can check what the research found.

One conversation, one framework

The framework menu next to the message box now changes that conversation only. It carries everything said so far across, and it picks the account the conversation runs on. New conversations start on the framework and account set in Settings.

The menu shows every framework with its logo, all at once, with no scrolling to find the one you want. There are thirteen now: Claude Code, OpenAI Codex, Gemini CLI, OpenCode, Kimi Code, Qwen Code, MiniMax Code, CodeBuddy, Grok Build, Cursor, Hermes Agent, and two new ones in this release, OpenClaw and NanoClaw.

Models and modes work the same way. A model or a mode you pick applies to that conversation only; the defaults for new ones are in Settings.

Moving a conversation, step by step

1. Open the conversation you want to move.
2. Click the framework menu next to the message box and pick another framework, say Codex instead of Claude Code, and the account it should run on.
3. If the agent is working, it stops that turn first, then moves the conversation with everything up to the stop.
4. Watch the line above the message box. It says where the conversation is going and how many messages it carries. Anything you type meanwhile waits and goes there once it has arrived.
5. Carry on. A short note in the conversation says what answers now and why.
6. To change where new conversations start, open Settings → Models & accounts and press Use on another configuration. The selected one is the default.

What travels with it

Switching framework leaves the conversation exactly as it was: your messages aren’t shown twice, and the conversation isn’t listed twice.

• The long ones too. The most recent part of a very long conversation goes along in full. Everything before it is kept, word for word, in a file the new framework is told about and reads when it needs something from earlier.
• What the agent did. The new framework is told which commands ran, which files were touched and what the latest steps returned, not only the tools’ names.
• Tesseract’s tools. From its first answer, the new framework can ask you a question, show a form or choices, keep its plan in the planner, start a helper and schedule work. Forms, choices and the plan checklist now work on Codex, OpenCode and Gemini CLI, not only Claude Code.
• Who said what. Under each reply you see the framework and model that wrote it, with its tokens, cost and time. Older replies keep their own after you switch.

One model menu, everywhere

Choosing a model is now the same control in the conversation’s menu, in Settings and in a configuration. Click it and the dropdown itself becomes the box you type in. The list starts with the six models you used last, then every model under its maker’s name, and typing narrows it by model or maker. Only models the selected framework can run are offered.

On the Tesseract Gateway, every model now comes from one record. The gateway serves each one that was seen to answer, about 270 today, and the menus are built from the same records and refresh every 10 minutes, so a new model no longer waits for a release. The Effort slider works on every framework too, and shows the levels the selected model really has.

Your own accounts, and your old conversations

If you add your own accounts in Settings → Models & accounts, an account that has used its allowance is now left alone until the time its provider gives, instead of being tried first on every message. The list shows each account’s session and weekly use, and you can turn an account off and back on.

Conversations you started outside Tesseract are in the list too. The conversations list and search now include what you began in the Codex app, OpenCode, Gemini CLI, Cursor, Grok, Kimi, Qwen Code, CodeBuddy, MiniMax and Hermes, next to your Claude Code ones. Opening one keeps a copy in Tesseract, and Continue carries it on with everything said so far.

Try it

Tesseract updates itself, or you can download 0.11.26 for Mac, Windows or Linux at tesseract.dev. Open two conversations, give each a different framework, and see which one you like for the job.

A few questions, then a plan

Click Make a presentation on Home and Tesseract asks you four things:

• what it is about,
• who will watch,
• how long it should be,
• and which look you like.

Tap one of the suggested answers, or type it in your own words, in any language. If you’d rather not choose, press Decide for me.

Your answers become a plan, and you get to change it before anything is built. Rename slides, put them in a different order, and trim the ones you don’t need. Then the slides are built while you watch, in a card beside the chat.

Start from what you already have

You don’t have to start from a blank page.

• From a document. Attach a Word file, a PowerPoint, a PDF or your notes, and the outline is written from it. Pictures you attach are placed once the slides exist.
• From one of your own. Pick one of your recent presentations under the first question. The new one keeps its colours, fonts and background and follows its slides in the same order, with new words.

Step by step

1. On Home, click Make a presentation.
2. If you have a document to start from, attach it. To follow an earlier presentation, pick it under the first question.
3. Answer the questions by tapping or typing, or press Decide for me.
4. Look over the plan. Rename, reorder or trim slides until it says what you want.
5. Watch the slides being built in the card beside the chat.
6. Ask for any change in the same chat. The slides update as the agent works.
7. Present it, or download it.

When it’s done

The finished slides sit in a card next to the conversation that made them. From there you can:

• present, or present with your notes;
• download as PowerPoint, PDF or a web page;
• change the look in one click. There are six looks, and every slide follows.

Headlines size themselves to fit, in every look. And Present now really takes the whole screen. It used to stay inside the Tesseract window, with the title bar and the tabs still showing.

Because the deck stays beside its chat, changes don’t mean starting over. Ask for a shorter version, a different order or a new slide, in the same conversation, and the slides follow.

Also in 0.11.26

A few other things you’ll notice in this release:

• Each conversation runs on its own framework. One conversation can run on Claude Code while another runs on Codex, at the same time. The framework menu next to the message box changes that conversation only, and carries everything said so far across.
• Work carries on after a restart. If Tesseract closes, crashes or restarts while the agent is working, the interrupted turn carries on when the app is back, with a note saying “Picked up where it left off”. This now works on every framework.
• Tesseract for Windows is signed. The Windows app, its installer and the Remote Desktop and Private Network apps are signed by Eightcell Labs, Inc, the company behind Tesseract, so Windows shows a verified publisher. An update is installed only when it carries the same signature.
• A notification that needs you opens the conversation. When a conversation you aren’t looking at is waiting for your approval, your input or a decision, a click on its notification, or its Open button, now opens it.
• Thinking is back in the chat. A Thinking switch sits right under the modes in the sliders menu. With it on, which it is unless you turn it off, the model’s reasoning shows in the conversation as it works, folded away with a click.
• Remote Desktop no longer flickers. The other computer’s screen stays on screen while you move the mouse over it. More on Remote Desktop.

Try it

Tesseract updates itself, or you can download 0.11.26 for Mac, Windows or Linux at tesseract.dev. Then click Make a presentation on Home and answer the first question.

Your computers, in one list

Sign in with the same account on each of your computers and they show up for one another. Open Remote Desktop and the Computers tab lists your machines, with whether each is online and which system it runs. Pick one to view and control it.

A computer keeps one identity in your account. It keeps the same Remote Desktop number whether or not anyone is signed in, and it keeps its place in your account even if Tesseract’s folder is deleted or the system is reinstalled.

For a computer that isn’t on your account, connect with its machine ID, a code or an email address instead. A machine ID comes with a password, for computers nobody is sitting at.

Connect, step by step

1. Install Tesseract on the computer you want to reach, sign in with your account, and leave it running. Updates wait for an idle moment and never install while a remote-desktop session is live.
2. On the computer in front of you, open Remote Desktop.
3. On the Computers tab, find the machine and press Connect.
4. For a computer outside your account, type its machine ID, code or email in the box under the list. With a machine ID, add its password. Then press Connect.
5. Work on it as if you were there.

What changed in 0.11.25

• It feels like sitting there. Controlling a Mac or Linux computer from afar now drags, double-clicks, right-clicks, holds Shift or ⌘ while clicking, and types in any language, accents, Cyrillic and emoji included. A key you let go of is let go on the other computer too.
• Start at login, on Windows and Linux. The Remote Desktop app’s Start at login pointed Windows at a temporary copy that was gone by the next sign-in, and did nothing on Linux. It now starts from the app you downloaded, quietly in the tray.
• Straight to your Windows PC. If you answered Windows’ “allow access” question the usual way, private networks only, Windows also shut out your own computers on your Tesseract private network, so every connection took the slower route through a relay. Your private network now gets through, and your answer still holds everywhere else.
• Honest about Wayland. On Linux under Wayland, clicking and typing from afar can’t work. Tesseract now says so and suggests the Xorg session, instead of reporting clicks that went nowhere.

Controlling from Windows got the same care a few days ago in 0.11.23: mouse buttons land where you click on a scaled display, and accented and non-Latin characters arrive as you typed them.

Getting through, wherever you are

Remote Desktop connects directly when the network allows, and through a relay when it doesn’t. Your Tesseract private network, which links your computers, got a run of fixes earlier this month:

• A computer behind a home or office router now discovers the address the internet sees it at and shares it with your other computers, so two computers on different networks can connect.
• A computer with both a newer and an older style of internet address offers both.
• If the private network’s helper stops and restarts, the connection comes back on its own.
• On your private network, you can reach a computer by its own name instead of its number. The names work only between your computers; nothing is published anywhere.

A black screen now comes with a reason. Three ways a connection could fail used to leave you looking at a black window in silence; each now says what happened, including the common case where the other computer has screen sharing turned off. And the one that connected, stayed black for twenty seconds and gave up is fixed.

Files, not just the screen

You don’t always need the whole screen. Your other computers are reachable from the file browser, from Explorer and from a machine’s own page, and browsing starts in that computer’s home folder. Images preview as you scroll, fetched only when you need them.

While you are viewing another computer, files you copy over there are offered to you here: drag them out or paste them. Only the list crosses the connection until you take them. Since 0.11.25, copying big files or whole folders through the Tesseract drive is several times faster and no longer stops halfway, and an interrupted copy never leaves a broken file. On Windows, with the private network set up, files up to 4 GB copy in File Explorer and huge folders open, from your next restart.

Try it

Download Tesseract for Mac, Windows and Linux at tesseract.dev, sign in on each of your computers, and open Remote Desktop. There is also a free Remote Desktop app of its own.

Computer control, on every framework

Computer control lets the agent use your computer much as you do: look at the screen, click and type. The rules are the same whichever framework the conversation runs on:

• Looking at the screen never asks.
• Clicking or typing asks you on screen first.

So you can hand a job to Codex or OpenCode that needs another app, and it still asks before it clicks or types.

It works off the Mac too. Until this release, when the assistant pressed ⌘C or ⌘V on a Windows or Linux computer, it pressed the Windows key instead. It now copies and pastes.

Settings → Computer control is the one place to switch it, for every framework at once. Screen readers now announce that switch by its name, “Computer control, switch, off”, instead of a bare “switch”, and the same goes for every other switch in Settings.

Turn it on, and ask from any app

1. Open Settings and go to Computer control.
2. Switch it on. The one switch covers every framework.
3. In a conversation, pick the framework you want from the framework menu. Picking another one now really switches to it, and the menu shows the framework that answers.
4. Ask for something that needs your screen or another app.
5. When the agent wants to click or type, it asks you on screen first.

You don’t have to be in Tesseract to start, either. The quick prompt opens Tesseract over whatever you are working in, with a picture of your screen attached. On a Mac it is ⌥Space. Off the Mac, the Computer control setting now names the real shortcut, Alt+Shift+Space.

Chat mode, for questions

Sometimes you only want to ask a question. Chat mode answers, reads your project and searches the web, and it never changes anything. Pick it from the mode button under the message box, or press Shift+Tab until it says Chat.

The mode menu is clearer too. The mode button names the mode you’re in; above, it says Full access. The model shows which framework’s default it is, and Advanced groups its settings with a line on what each check does. That leaves three modes:

• Chat answers and searches, and changes nothing.
• Plan reads and proposes, and changes nothing.
• Full access does the work.

Keep going when I’m away

“Keep going when I’m away” used to be a mode of its own, and it read like a step beyond Full access. It is now a switch under Full access, and it is more careful:

• When the agent asks you something, it waits two minutes for you before answering, and counts down on the card.
• It never approves a step flagged as risky. It skips that step and tells you.
• When you come back, the notification says how many decisions were made for you and opens straight on the report. Clicking any conversation’s notification now opens that conversation.

Also in 0.11.25

• It says what it’s waiting on. When a run goes quiet, the line above the message box says why, whether it is your answer, summing up older messages or a long command, and for how long.
• Edit a request that can’t be done. When a run stops because the request can’t be done as written, the card has an Edit request button that puts your message back in the box to change and send.
• Closing a terminal stops everything in it. Background jobs, dev servers and nohup commands no longer keep running after you close their terminal or quit Tesseract, on every system.
• It says why a framework can’t install. Instead of an error code, you see what is wrong: Node.js missing or too old, the network, the disk, or a permission.
• Remote control that feels like sitting there. Controlling a Mac or Linux computer from afar now drags, double-clicks, right-clicks and types in any language. More on that on the Remote Desktop page.

Try it

Tesseract updates itself, or you can download 0.11.25 for Mac, Windows or Linux at tesseract.dev. Switch on Settings → Computer control, and try Chat mode for your next question.

The first run

The first time you open Tesseract, setup is one guided flow in one place, with a step indicator: the terms, the permissions the app needs, and the framework that does the work.

A framework is the AI agent that reads, writes and runs things for you. Tesseract runs one installed on your computer rather than bundling its own, so you can use the accounts you already pay for. From today, new installs start on OpenCode, because it works out of the box with the models included in Tesseract.

The other frameworks are hidden for now, not removed. If you installed one before, nothing was touched, and if you had chosen a different framework, Tesseract still remembers it.

Setup also got sturdier over the last few releases. If an install fails, the screen hands you the commands to run by hand. And if the check for what is already installed goes wrong, the setup screen no longer sits on “Installing…” with no way past it.

Home: start on the left, come back on the right

Home is where you start. What you start is on the left: the Ask anything box and a set of illustrated starters, such as Build a website, Make a presentation and Do deep research. Some carry a New badge, like Make a video and Create a font. What you come back to is on the right, starting with your recent projects. New conversations open on the same starters.

A sentence is enough. Type what you want, or start from one of the ready ideas if you’d rather not type at all.

The menu under the message box sets how much the agent may do on its own. There are three ways to run:

• Plan reads and proposes, and changes nothing.
• Full access lets it do the work.
• Unattended is Full access that also answers the agent’s own questions, for work you leave running.

The finer switches are in an Advanced panel, and the defaults for new conversations live in Settings.

Signing in

You don’t need an account to start. If you have one, sign in with the account button. It used to be only on the Home tab; from today it is on every Settings page too, so you no longer have to go back to Home to sign in.

A few things around your account changed in this release:

• Your picture. Sign in with Google and Tesseract picks up your Google picture. You can upload your own in Settings → Account, or remove it and go back to your initials.
• Your password. Settings → Account can change your password. If you have only ever signed in with Google, it offers to set one, so you can sign in with your email too. Changing it signs your other computers out and keeps you signed in on this one.
• The sign-in tab. After signing in with Google, the browser tab used to be a bare line of text. It is now a proper page that tells you what happened and closes itself after a few seconds, with a button in case your browser won’t let it. If you declined Google’s permission screen, it says what went wrong instead of “Signed in”.

Once you are signed in, this computer is part of your account. Your other signed-in computers appear in My Computers, and this one keeps its place even if you delete Tesseract’s folder; after a system reinstall or a new disk, it is recognised rather than added twice.

Your first ten minutes, step by step

1. Download Tesseract for Mac, Windows or Linux and open it.
2. Go through setup: the terms, the permissions, and the framework. On a new install, that is OpenCode.
3. On Home, type what you want in Ask anything, or click a starter.
4. In the menu under the message box, pick Plan to see what the agent proposes before anything changes, or Full access to let it work.
5. If you have an account, sign in with the account button on Home or on any Settings page.
6. Stuck? Press ⌘/ and ask the guide out loud.

Two settings worth knowing

Updates are yours to control. You can turn automatic updates off in Settings, or by right-clicking the update button in the bottom-right corner. Tesseract doesn’t go quiet when you do: it still tells you a new version is waiting. It just won’t download or install one until you say so.

The guide speaks your language. The narrator, the guide on ⌘/ and the assistant on the message box speak any of the 99 languages the voice service supports. Pick one in Settings, ask out loud, or choose it during a walkthrough, and all three follow. The guide is a voice only: there is no chat window beside it, and opening it ends a walkthrough that was running instead of talking over it.

Try it

Download Tesseract 0.11.22 for Mac, Windows or Linux at tesseract.dev, open it, and ask for the first thing on your list.