Changelog

Type /goal and what you want. The team works in rounds on its own, checks its progress after each one, and stops at the limits you set. A bar shows how far along it is, and a result line says plainly whether the goal was met.

The Start a goal screen shows, under each check, exactly what it will look for, so a loosely worded sentence is not mistaken for a check that can pass. If the project is set to Full machine it says so up front, because a goal runs with nobody to approve anything and never uses that profile.

Run a thread in a separate copy of your project. When it is done, look over the result and apply it back, with a clear dialog if some files clash. Delete the copy if you do not like it.

Set a task to run at a time or on a repeat while Maxtkn is open. Results land in an inbox. You can also copy a ready-made command to run a task from your computer’s own scheduler.

Run the same team from a terminal, with --goal for an unattended goal, --worktree for an isolated copy and --check to verify a project. A GitHub Action runs it in your own workflows.

When agents change files, Maxtkn runs your project’s own tests, type check, lint or build. A failure goes back to the team to fix, and the summary says what was verified or why something could not be.

Turn on Use a browser (or type /browser) and the team can open your app, click through it, read its console and take screenshots. Only your project, localhost and sites you allow can be opened.

From the desktop window the hidden browser used to refuse the connection, so Use a browser always failed with “Could not connect to the browser”. It now allows the app’s own window, and no other page.

A check that opens the app failed with “not a valid address” in a project with no dev server, such as a single index.html. It now opens the page from the project folder.

Write what should be true, such as “the Add button adds a task”. See the plan in plain steps, save it with the project and run it again any time from the new Checks tab.

After a run, a reviewer reads the changes and lists findings. Jump to the change, ask for a fix with one click, or dismiss a finding and it stays dismissed.

Buttons under changes, errors, failed steps and finished runs explain them in everyday language, with a Simple, Detailed or Code switch. A finished run also says what changed by itself.

Maxtkn saves a checkpoint before each message and approved command. Pick any point and restore your files to it, with a preview of what will change and what cannot be undone.

The review pane now works change by change. Keep or reject each one, leave comments that go to the team as one follow-up, compare side by side and use the keyboard.

A people icon in the Files tab lists every file with who last changed it, when and why. A file’s page shows its whole history, with a link back to the conversation and to the saved copy from just before. It is built from what the team and you did, nothing is guessed.

See what changed, choose what to include, save a version, switch branches, upload and download, and read the history, all with plain-language messages instead of git jargon.

Plans, walkthroughs, reviews and notes get their own tab with a version picker and a link card in the thread.

Select words in a plan or walkthrough, add a comment and send your comments to the team. They write the next version of the same document, and you can compare any two versions word by word.

The Preview tab can start your project’s dev script after asking, show its log, restart it and switch between phone, tablet and desktop sizes.

Point at a button, heading or box in the preview, then say what to change. The agent is told which element it is, where it lives in your files and which style rule paints it. It works on dev-server pages too.

Draw boxes and arrows on an attached screenshot or mockup so the team sees exactly what you mean.

A button in the Preview tab takes a frozen picture of the page and opens it in the markup dialog. Draw on it, and the marked picture lands in your message box with your marks described in words.

A new button in the Preview tab opens the page in a separate window you can drag to any size. It has phone, tablet, laptop and wide presets and a live width and height readout, and it reloads by itself when the files change. Press the button again and the same window comes forward.

Turn on pointing inside the preview window and click anything on the page. It shows up as a chip in your message box, and a Describe the change button takes you straight there. It works for your own page and for dev-server pages.

Settings → General → Appearance: choose Light, Dark or Match my device, turn on high contrast for brighter text and clearer borders, or reduce motion. The editor and terminal follow, and a screen reader is told when the team starts, needs you or finishes.

Hover hints are now soft floating cards with a heading, a short explanation and the key to press. The features that most need explaining, like pointing at the page, thinking effort, isolated copies, steer or queue, keep and undo, and each workspace tab, also play a few seconds of an animated mock-up with a cursor that clicks through it. With reduced motion switched on, you see the last frame instead.

The Terminal tab is now a full terminal with tabs, your own shell, colours and Ctrl+C, opened in the project folder. Agents can read what it shows but cannot type into it.

Every file and folder has a menu to rename, move, duplicate, copy its path, show it in your file manager or delete it, and folders can make a new file or folder inside. Names are checked as you type and nothing is overwritten. Changes to text files are recorded as yours, so a deleted file can be brought back from the Changes tab, and the delete says first when it cannot be.

Ctrl or Shift-click picks files and folders. Move and Delete work on all of them, you can drag them onto a folder, and one button closes or opens every folder at once.

In a window narrower than about 1,100 px the conversation used to be squeezed to a sliver. Now the sidebar steps aside, the work panel gives way in very narrow windows and its tabs scroll. The message box also re-fits when you resize the window.

A thread opens with its newest 150 steps and a Show earlier steps button, and a finished file write no longer stores the file twice.

Right-clicking no longer shows the web view’s own menu, and the developer keys, view source and reload are switched off. Text fields keep cut, copy and paste.

With a dialog open, Tab no longer wanders off to the page behind it.

Off by default. Switch it on in Settings to get an Edit button on files, changes and the review pane. Your edit is saved only when you press Save, shows up in the thread as yours and can be undone.

Ctrl+K finds actions, threads, files, commands and skills by a few letters. Ctrl+P opens a file by name and Ctrl+Shift+F searches the text of the project.

Type @ in the message box to point at a file, folder, other thread, skill, agent or document. Protected files are never shared.

Open a file at the right line in the editor you already use, straight from the Files list or a change card.

Agent messages, their thinking and the run summary now show headings, lists and code properly instead of raw symbols.

The message box stays usable during a run. Steer the team right away, or queue your next message to send when it finishes.

For work that changes structure, the Architect shows a design before anything is built: a short summary, diagrams (flowcharts, sequences, data models) and the decisions behind it. Press Approve and build, or Ask for changes with a note. It is skipped for small edits, and the approval is not asked in Full access or in runs nobody watches. Settings → General → Design review chooses whether the Architect decides, always designs or never does, and any mermaid block in a plan or document is now drawn as a diagram.

Ask for a plan first. The team can ask clarifying questions, and nothing starts until you approve.

Branch a thread from a message to try another direction, ask a side question, or edit a message and resend it.

Search every conversation, pin the ones you need, archive the rest and save a thread as a page.

Ask a question with no project at all. Quick chats are read-only, and you can turn one into a project later.

Type /compact to summarize the earlier runs of a thread. The summary is labelled and can be undone, and a meter shows how full the conversation is.

Add images and documents with the + menu, drag and drop, or paste. They are sent to models that can read them.

Every thread, grouped by what needs you, with approve and deny right there and badges that show where attention is needed.

Get a system notification when a run needs approval, asks a question, finishes, fails or pauses. Mute any project.

Already pay for Claude Code or Codex? Pick one of their models and Maxtkn runs your signed-in tool for you. This is an extra option next to API keys.

Presets for OpenRouter, Gemini, Mistral, Groq, Azure OpenAI, Ollama and LM Studio. You can see what each model supports before you pick it. OpenRouter’s list is no longer limited to coding models, so MiniMax, more Qwen models and newly released ones show up.

The model menu, the chip that shows your choice and the lists in Settings show each model’s maker or its provider. Under each name is just the provider, and OpenRouter names no longer start with the vendor.

Settings → Models → Models in the chat menu lets you tick the models you want in the message box. Only models you can use now are listed. Until you choose, the menu shows the newest Claude of each tier plus the models you added.

A model that is being retired says so in the model menu and the schedule editor. The team designer says hand-offs are guidance, and a run that skipped one lists it. A lead that stops mid-work is asked once to carry on.

Choose how hard a model thinks, per run or per agent, and see its thinking in the timeline.

If a provider fails, Maxtkn retries, switches to your fallback model, or pauses the run with a Resume button instead of crashing the thread. A provider that goes quiet for two minutes, or a connection that drops halfway through an answer, is retried too, and an answer cut off for lack of room is continued where it stopped.

Every model turn used to be limited to 8,192 tokens, so a whole page written in one go was cut off halfway and nothing was saved. A turn now gets the room its model really has, and a cut-off write is never run as an empty one: the agent is told to split the work into smaller pieces.

Fetch models and Test connection now say in words that a provider could not be reached, instead of showing a raw error.

The message box says “About 40 cents · ~6 min” before you send, based on your own recent runs or a labelled guess. A finished run says how long it took and what it cost.

Settings → Usage is now a dashboard. A big spend figure says how it compares with yesterday or the period before, with a bar chart by hour or day. Below are calls, tokens, how much input came from the cache and a typical run, then where it went by model, project or day. Budgets show as plan meters with what is left and when they renew, and the numbers export as CSV.

Every command is checked before it runs. Dangerous ones are blocked, uncertain ones ask first, and the reason is shown on the approval.

Tick Always allow on an approval and Maxtkn remembers the rule. Review or remove rules in Settings.

Off by default. Settings → Safety → Run commands in isolation runs a project’s commands in a Docker or Podman container, or a bubblewrap sandbox in WSL. Only the project folder can change, and there is no network unless you allow it. If it cannot start, nothing runs.

Pick what agents may do on their own, keep secrets such as .env files off limits and allow only the web domains you trust.

Hooks and other runnable settings found in a folder stay inactive until you decide to trust them.

API keys are now kept in the system keychain instead of in the app’s database, and secrets are hidden from logs and model output.

Build a team visually. Pick agents, see them as cards, make one the lead and say who hands work to whom and when. The lead is told that order. The designer says hand-offs are guidance, and a run that skipped one lists it. Start from a template or share the team as a file. Agents and teams also shows which of your teams have really worked for you, with their success rate, usual time and cost.

Customize has an Import button that copies Cursor rules, Claude Code skills and commands, MCP servers, hooks and Copilot instructions into Maxtkn. You see the exact text first, programs wait for your tick, keys and passwords are never copied, and the original files are left alone.

Add agents, teams, rules, skills, commands, MCP servers and hooks in one step from a folder, a git address, a file or pasted text. Anything that could run a program or reach a website stays off until you tick it, and uninstalling removes exactly what the bundle added.

Agents, teams, rules, skills, commands, MCP servers and hooks in one place, showing where each one comes from and a switch for each.

Maxtkn reads your AGENTS.md and rule files, applies them when they fit, and shows which ones it used. One click creates an AGENTS.md.

Add skills and your own commands, and use them with a slash or a $name.

Connect MCP tools, with approval per tool. They stop when the run or the app ends.

Run your own scripts at key moments of a run. Hooks only run in trusted projects.

Agents can call in short-lived helpers and hand work to the background, with limits on how many and how deep.

A thread’s menu has Publish the site. It shows exactly which files would go, and which are left out on purpose: secrets, history, packages, hidden files and anything whose text looks like a key. Then it saves a zip or, with your own Netlify token kept in the system keychain, puts the site on the web after a last confirmation. The thread records it.

Save a thread as an audit record (JSON or Markdown): who did what and when, with each step linked to the one before it so a changed file can be found, and an optional seal key. A small script checks a record without the app. Settings → General can delete archived threads after 30 days to a year, and is off by default.

The scripted offline Demo is gone. Connect a model in Settings (an API key, an endpoint, or your Claude Code or Codex plan) and the first-run guide walks you there. The empty chat screen no longer shows ready-made idea buttons.

A short guide for new users and five project templates that work offline.

A per-user setup that needs no administrator rights, and an MSI for managed machines.

Settings has Check for updates with a stable or beta channel. Updates are verified against a signature before anything is installed.

Fixed a case where the window could turn blank.

The app was called Open IDE. The first launch after the change carries over your threads, settings, saved API keys and checkpoints, and each project’s settings folder. Isolated copies made before the change, and tasks already added to the operating system’s scheduler, are not carried over.

Each project is a folder on your computer, and each thread is one task.

A lead plans the work and hands tasks to specialists.

Try the whole app with no API key.

A panel beside the conversation shows your files, who changed what, a terminal and a live preview.