Early access · Windows
Agent teams plan, build, test and review real projects on your machine. You approve what matters.
Works with the models you already use
The problem
If you describe software in plain language, you deserve to understand and steer what gets built, not approve diffs you cannot judge.
You get a wall of diffs and are asked to judge it. If you are not a programmer, you are approving blind.
Planning, building, testing and reviewing all happen in one long chat, so you cannot tell who did what, or why.
Changes land in your files with no clear owner and no easy way back, so every run feels like a risk.
Features
Everything an agent does is visible, attributed and reversible, so you can stay in charge without becoming a programmer.
A lead plans, specialists build, and independent tasks run at once. Bigger work starts with a design you approve first.
Independent tasks run at the same time.
Choose what agents may do alone. Risky commands and secrets like .env files always ask first.
Click a mode to try it.
See who wrote each change and what changed. Keep or reject single hunks, or rewind the whole project.
.card {
− padding: 8px;
+ padding: 24px;
+ border-radius: 16px;
}
A live preview sits beside the chat, your files, a code editor and a real terminal. Draw on the page to show what to change.
Task Tracker
Plan the launch
Write the post
Book the demo
See what changed, save a version, switch branches and read the history, all in plain language.
Add the task list and its styling
Each project is a folder, each thread a task. Run several at once, search them all, or fork from any message.
Each thread is one task. Run several at once.
The message box shows the likely cost and time before you send, and a finished run says what it took.
Add a dark mode switch to the settings page
About 40 cents · ~6 min
The estimate comes from your own recent runs.
Save a thread as a record of who did what and when. A script can check that the record was not changed.
Agents only touch your project folder, and commands are checked first. Keys stay in Windows Credential Manager and history stays on your device.
Documents/Maxtkn/task-tracker
agent ▸ write_file index.html
Written inside the project
agent ▸ write_file ../secrets.txt
Path must stay inside the project folder.
Checked on every file operation.
Checks
You should not have to read code to know if something works. Maxtkn runs the checks, looks at the running app, and shows you the evidence.
After the team edits, Maxtkn runs your tests and build. Failures go back to the team to fix.
An agent opens your running app in a real browser, reads the page, clicks, types and takes screenshots.
Write "the Add button adds a task" and get a saved check you can run again.
The final walkthrough lists what the checks found, not the model grading its own work.
A reviewer lists findings you can jump to. Fix one with a click, or dismiss it.
Ask for any change, error or command in plain words, at the depth you choose.
Evidence
ExampleWritten from the results of the checks. A check that could not run is listed as not run.
How it works
Watch a run from the first click to a finished app. Connect a model once and you can run the same flow yourself in minutes.
Illustration of the app. Hover to pause, or pick a step.
Agent teams
A lead plans and delegates while specialists work in parallel. Start from a template, or design your own team.
1 agent
One Builder agent works directly. Fastest and cheapest for small changes.
3 agents
A Lead plans and delegates to a Builder, then a Reviewer checks the result.
6 agents
Lead, Architect, Designer, Builder, Tester and Reviewer. Specialists run in parallel.
Terminals
Start up to twenty sessions in one window, each with its own tool. Every session works in its own copy of your project, so nothing collides.
Terminals
New sessionsStart one to twenty sessions. Each gets its own copy of the project and its own branch.
An example. The sessions and what they print are made up.
Each session gets its own git branch and worktree, so two sessions never overwrite each other.
Pick how many and which tool: Claude Code, Codex, Cursor, OpenCode or Gemini CLI. A plain shell works too.
Each session shows if it is working, waiting for you or stopped. The left rail counts the ones that need you.
Send one line to the sessions you pick. Start, resume or stop them all from one toolbar.
The header shows the memory your sessions use, and each session shows its own.
Claude Code
Your Claude plan
claude
Codex
Your ChatGPT plan
codex
Cursor
Your Cursor plan
cursor-agent
OpenCode
The model keys you gave it
opencode
Gemini CLI
Your Google Gemini plan
gemini
Plain shell
Nothing
a normal terminal
Your own
Whatever it runs
any command you add
Claude Code is the most tested. The other four show a “not run yet” note when you choose them in the app.
Sessions run the tools you have installed and signed in to, on your own computer; Maxtkn signs in to none of them. Open Terminals from the left rail, from the Terminal tab’s Many terminals button, or with Ctrl+K. Read how it works.
Control
Choose an access mode per thread and a permission level per project. Whatever you pick, agents stay inside the project folder, and every change can be undone.
Every file change and every command waits for your approval.
When you want to see everything.
Agents create and edit files freely. Deleting files and running commands still need your approval.
The balanced default.
Files and commands run without asking. Dangerous commands still stop and ask.
When you trust the task and the team.
The access mode says when to ask. The permission level says what is possible at all.
Autopilot
Hand over a goal, a schedule or a pipeline, and come back to a report. Unwatched work runs under stricter rules.
Say "make the tests pass". The team keeps working until a real check confirms it, or a time, round or money limit stops it.
Hourly to monthly tasks run while Maxtkn is open and report to an inbox. You allow each one first.
A thread can work in its own git branch and copy. Nothing reaches your project until you click Apply.
Run a request from a terminal or CI job. The exit code says if it finished, failed a check or needed an approval.
When nobody is watching, anything that would ask for approval is denied and listed in the report. Runs never use Full access, and always have a time limit and, for priced models, a money limit.
$ maxtkn run --verify "add a dark mode switch to the settings page"
Exit code
Command line
The same team and the same safety rules, without the window. Here is every command-line tool and where each one stands today.
Run a whole request from a terminal or a script. Choose the team and model, check the result with --verify or --check, keep going until a goal is met with --goal, or work in an isolated copy with --worktree.
Built from the project source. Not published as a package yet.
Run a request in a pipeline. It uploads the event log, puts the walkthrough in the step summary and fails the step with the command line’s exit code.
In the project source. Not on the Marketplace yet.
Copy a ready-made Windows schtasks, macOS launchd or Linux crontab command that runs a scheduled task even when Maxtkn is closed. It never contains a key.
In the app, under Scheduled tasks.
Start Claude Code, Codex, Cursor, OpenCode or Gemini CLI in many sessions at once, each in its own copy of your project.
In the app, under Terminals.
Your setup
Rules, skills and tools from other AI coding tools carry over, and each one stays under the same approval rules.
Reads AGENTS.md and .cursor rules as they are, and applies them always, to certain files, or when you @name them.
SKILL.md folders made for other tools work as they are. Switch any skill off per project.
Connect MCP servers over stdio or HTTP. Each tool asks for approval first.
Copy rules, skills, commands, MCP servers and hooks from Cursor, Claude Code and Copilot. You see the exact text first, and keys are never copied.
Pricing
Maxtkn for Windows is in beta. No subscription: one payment, the same full app, and the only difference is the support you get.
Community
$29Regular price $79one time
Launch offerCode LAUNCH. Enter the code at checkout.
The full app, with help from the docs and other users.
Supported
With support$199one time
The full app, and a person to write to when something breaks.
Maxtkn for Windows is a beta. Model usage is billed by the provider whose key you use and is not part of the price. Licenses are not on sale yet: join the early-access list and we will send you the launch code when they are. Once they are, you can ask for a full refund within 7 days.
Install
Download it, allow it once, and open it. No technical knowledge needed.
Press the button and the installer lands in your Downloads folder. It is one file, and it works on Windows 10 and 11. You do not need an administrator password.
On a Mac or Linux? Join the list and we will write when your build is ready.
Maxtkn-Setup.exe
Done · in your Downloads folder
Your browser may say the file “isn’t commonly downloaded”. That is the same beta notice. Choose Keep to save it.
Double-click the file. Windows will stop you with a blue screen, because Maxtkn is a beta and Windows does not know it yet. Choose More info, then Run anyway. Then follow the installer.
Windows protected your PC
Microsoft Defender SmartScreen prevented an unrecognized app from starting. Running this app might put your PC at risk.
More info
Windows protected your PC
Microsoft Defender SmartScreen prevented an unrecognized app from starting. Running this app might put your PC at risk.
When the installer finishes, Maxtkn opens. A short welcome takes you to Settings, where you connect a model: an API key, a local model, or your Claude Code or Codex plan.
Maxtkn is ready
Find it in the Start menu, or leave “open Maxtkn” ticked at the end.
Maxtkn is in beta, and it is not certified by Microsoft yet. Windows shows this screen for every app it does not recognise, so it is normal here. It does not mean something is wrong with the file. Only run an installer you downloaded from this website.
Windows only for now. To update, download the newer installer and run it over the old one; your projects and settings are kept.
Early access
Maxtkn is a beta for Windows, and you can download it above. For macOS or Linux, leave your email and we will write once, when a build is ready.
FAQ
No. Maxtkn is built for people who describe what they want and review results. The code is still one click away if you are curious, but you never have to read it.
After a run, Maxtkn runs your project checks and any acceptance checks you wrote in plain sentences, can open the app in a browser, and lists the evidence in the walkthrough. A check that could not run is reported as not run, never as a pass.
Anthropic Claude, any OpenAI-compatible API (OpenAI and similar), presets for OpenRouter, Gemini, Mistral, Groq and Azure OpenAI, and local servers such as Ollama or LM Studio. You add your own key in Settings, or use your Claude Code or Codex plan.
The app itself runs on your computer, but models need a connection unless you point Maxtkn at a local model server such as Ollama or LM Studio.
Yes, within limits. Goals, scheduled tasks and the command line run without you, with a time limit and a money limit for priced models. They never wait for an approval and never grant one, so anything risky is denied and listed in the report. Scheduled tasks run while Maxtkn is open.
Yes. Maxtkn reads AGENTS.md and .cursor rules, SKILL.md skill folders, slash commands from .claude and .cursor, and MCP servers. A folder that brings its own hooks or tools asks for your trust once, and lists exactly what would run.
On your computer. Projects are ordinary folders you choose, your keys are kept in the Windows Credential Manager, and your history and settings are stored in a local database. Model requests go directly from the app to the provider you configured.
Agents are limited to the project folder, paths that escape it are rejected, and commands are checked before they run. Access modes and permission levels decide what needs your approval, and risky commands always ask first. Every file change can be reverted, and you can rewind the project. By default this is a guard, not a virtual machine: you can switch on running commands in a container or WSL sandbox, and should keep Ask first for work you are unsure about.
The desktop app is built and tested on Windows today. macOS and Linux builds are not available yet. Join the early-access list and choose your platform to hear when they are.
Maxtkn is a one-time payment with no subscription: Community is $79 with community support, and Supported is $199 with direct email support. There is also a launch offer at $29 with a code. Licenses are not on sale yet, so join the early-access list and we will write when they are. Maxtkn does not resell model access. Model usage is billed by the provider whose key you use, and local models have no usage cost. Maxtkn shows an estimated cost for each run and lets you set limits.
Yes. If you buy a license and do not like Maxtkn, ask for a refund within 7 days of paying and you get the full amount back, with no reason needed. Refunds are not automatic, so you have to ask. The Terms page says how.
Yes. The Terminals view starts one to twenty sessions in the project, each in its own copy of it on its own branch, so they can build different features without overwriting each other. Each session can run Claude Code, Codex, Cursor, OpenCode, Gemini CLI, a plain shell or a command you add. Each tool must be installed and signed in on your computer; Maxtkn signs in to none of them. Sessions use your computer’s memory, so the header shows how much the running ones use. This feature does not need Docker.
Download the Windows installer from the early-access section when a build is ready, compare its SHA-256 with the card, and run it. It installs for your user without administrator rights. Early-access installers are not code-signed yet, so Windows may warn you first. The documentation has the full steps, including updating and uninstalling.
Yes. maxtkn run does a whole request from a terminal or a CI job, with the same safety rules and an exit code that says how it went. It is built from the project source today and is not published as a package or included in the installer yet.
Yes. The message box shows an estimate such as "About 40 cents, around 6 minutes" worked out from your own recent runs, or a labelled guess when there is no history yet. A finished run says what it took, and day and month budgets warn at 80% and stop at 100%.
Tell it what you want done. You review the result, and you can undo any of it.
Beta · Windows 10 and 11 · Installs for your user only
Your goal
ExampleAdd a dark mode switch to Settings, and make sure the tests pass.
Ready for your review
See what changedUndo