CrewKite runs the coding agents you already pay for as one crew on your own computer, and keeps the work going when one of them reaches its limit.
Side questions: Local gpt-oss-20b. Subagents on for every role.
Each provider's own client does the work, on your sign-in and your plan. CrewKite runs them as one crew, keeps the work moving at a limit, and keeps your checkout as it is until you accept a change.
Plan with Opus, build with Sol, fix with Argon and review with Astra in one thread, each on your own subscription.
When a provider runs out of included allowance, the work moves to the next connected one with its messages and edited files.
The rail shows each provider's usage windows and when they reset, read on a timer from your own plans.
The kite meter shows what the next message will cost against each window before you send it.
Allow credit spending starts off for OpenAI and Anthropic, so once a plan runs out, the next turn moves to another provider before it spends a credit.
Ask a quick question on a local model while the builder keeps working.
Every thread works in its own copy of your repository. Your checkout stays as it is until you accept a change.
Flight mode decides how much the crew does on its own, and Permissions decides which commands it may run without asking.
Run tests runs your project's test command in the thread's copy and shows the exit status and output, with no model involved and no credits spent.
CrewKite runs each provider's own client on your own plan. It adds no markup and holds no API key, and its OpenAI client has its own home, so nothing shows up in your own OpenAI app.
Load Qwen, gpt-oss, Gemma or any chat model your hardware can run in LM Studio, and it joins the crew with no usage limits. CrewKite runs the agent loop for it and sizes every prompt to the model's context.
The same crews and threads from a CLI.
Each provider runs through its own client, signed in with your own plan. CrewKite reads each plan's allowance and shows it in the sidebar. A thread can move between providers and keep its messages and edited files.
Antigravity, Google's coding client, on your Google account
As of 31 October 2026, Google gives accounts without a paid plan a quota that refreshes weekly. Google AI Pro refreshes every five hours until a weekly limit, and Google AI Ultra has the highest quota. Limits follow available capacity.
Codex, OpenAI's coding client, on your OpenAI plan
As of 31 October 2026, OpenAI says its coding client is included in its Free, Go, Plus, Pro, Business, Edu and Enterprise plans. Usage limits depend on the plan, the model and the task.
CrewKite runs this client in its own home with its own sign-in, so nothing it does shows in your own OpenAI app.
Claude Code, Anthropic's coding client, on your Anthropic plan
As of 31 October 2026, Anthropic includes its coding client in its Pro, Max, Team and Enterprise plans and leaves it out of Free. Pro and Max have a five-hour and a weekly limit. Ultra (formerly Max 20x) has a weekly limit only for verified accounts, as of November 2026; a newly opened Ultra account keeps a five-hour limit until Anthropic verifies it.
LM Studio, on this computer
As of 31 October 2026, LM Studio is free to use at home and at work. Speed and context follow your hardware.
Load any chat model LM Studio can run on your machine, such as Qwen, gpt-oss or Gemma, and pick it in CrewKite. CrewKite runs the agent loop itself. It gives the model tools to read, search, edit and run commands in the thread's copy, and fits every prompt to the context the model reports. Models without native tool calling work too. Connect works with LM Studio's authentication on or off, and Local reconnects by itself when LM Studio restarts.
G, O, A, L. Yes, we noticed.
On Linux, OpenAI, Anthropic and Local run as a preview. Google is not available on Linux yet.
Each provider connects once on this computer with an explicit Connect, then reconnects on its own at every later launch. Local also reconnects by itself when LM Studio restarts. The welcome screen walks through each one, and Settings keeps the same setup.
Google's published FAQ says third-party access through its coding client login breaks its terms. CrewKite has no exception. Google stays off until you read a short summary in Settings and accept the risk. Use it only if that trade is yours to make. The dated links are in the docs.
A crew is an orchestrator, a builder, a fixer, a reviewer and a challenger, plus the model that answers side questions. Each role runs on any family you have connected, with its own model and effort, and each can start subagents.
Side questions: Local gpt-oss-20b.
Subagents on for every role. Flight mode Balanced.
In the Full crew preset the orchestrator reads the request and proposes up to 8 slices. The controller checks the plan and assigns each slice to the crew's builder, or to its fixer for a hard slice, and nothing runs until you press Run plan.
The built-in presets are Solo, Tandem and Full crew. Save your own crews beside them, and set each role's model and effort.
A crew moves work to the next connected provider when one runs out of included allowance, and the rail shows every window as it goes.
When Anthropic reaches its limit, the thread continues on OpenAI with every message and edited file, and the next reply picks up where the last one stopped.
The rail shows the five-hour and weekly windows each provider reports, or the weekly and model windows for a plan with no five-hour limit. Google reports five-hour and weekly windows, OpenAI a weekly window, Anthropic the windows its plan has, and Local has no limits. CrewKite reads usage every 90 seconds for Google and OpenAI and every five minutes for Anthropic.
Gauge readings in these screens are samples from a test account.
Allow credit spending is a switch for OpenAI and for Anthropic, and it starts off. With it off, once a plan's included usage runs out, CrewKite starts no new turn on that provider: a crew step runs on the next provider, and a thread moves to the next one with allowance left. The rail shows "Credits paused" so you can see why. OpenAI spends purchased credits on its own once a plan's included usage runs out, and as of November 2026 I found no setting in its own apps or CLI that turns that off. A turn that is already running finishes, and the switch can never turn credits on. The provider's own setting still decides that.
The kite meter beside Send shows what the next message will use against the model's context window, separate from your plan's allowance. Here a side question runs on gpt-oss-20b in LM Studio beside the thread, and the builder keeps working.
Every thread works in its own copy of your repository. Read the changes, keep or drop each file, then accept them to a new local branch. Your checkout stays as it is.
Flight mode decides how much the crew does on its own, and Permissions decides which commands it may run without asking.
Run tests runs your project's test command in the thread's copy and shows the exit status and output, with no model involved and no credits spent.
Every thread works in its own copy of your repository. Run tests runs the test command CrewKite found in your project, such as the test script in package.json, pytest or cargo test, in that copy whenever you like. It asks before the first run unless you have allowed the command for the thread, shows exactly which command runs, and keeps the exit status and the output beside the thread. It never starts a model, so it costs nothing against any plan. Crews still run the checks you switch on after each fix and after the last slice of a plan.
I build products for underserved niche markets, and I start each one by talking with the people who have the problem. Along the way I kept running into problems in my own workflow that no tool solved.
I work across several AI subscriptions and a local model, because each one is strongest at something different: backend work, frontend and design, or catching defects in review. Using them together on one repository meant separate apps and terminals, watching every usage meter, and writing handoffs by hand. Each new agent reread the whole context, and the style of the work drifted with every switch.
CrewKite is the controller I wanted: every provider in one thread, held to how I want my work done. Its review methods draw on open-source projects like multi, ADHD and Ponytail, credited in the docs, so you get those ideas without having to find them first.
CrewKite is free and open source. Other people's open-source work has helped me a great deal, and this is my turn to give back. I am measuring efficiency with a benchmark now, and I will not claim gains until the numbers are in. Until then, what I want most is your feedback. My best product decisions have come from real people telling me what they need. Try it, and tell me where it falls short.
Raymond J. Kraft
Every thread works in its own copy. Results land as new local branches, and only when you accept them.
No telemetry, no account, no server of its own.
Tool output, exports and diagnostics go through redaction before they are shown or saved.
CrewKite never reads or stores a provider credential and never falls back to paid API billing.
The installers include the CLI. It talks to the same controller as the desktop, so keep the desktop open. Slash commands such as /model, /crew and /ask, file mentions with @, and -p for scripts.
Linux
crewkite thread list --jsonmacOS
/Applications/CrewKite.app/Contents/Resources/bin/crewkite thread list --jsonTry the CLI walks through it, and the CLI guide lists every key and command.
Configure a design system once and CrewKite enforces it on every provider.
Follow a pull request's CI from the thread, through your own GitHub sign-in.
Version 0.1.0 is the first public release. I build CrewKite in the open, every decision has an architecture decision record, and an extension seam lets separately licensed modules plug in.
A desktop workspace for coding with the agents you already pay for. It runs the Google, OpenAI and Anthropic clients you have signed in, plus a local model through LM Studio, as one crew on your own computer. A thread can move to another provider mid task and keep its messages and edited files.
CrewKite adds no plan gate. Each family works on whatever plan its provider allows, and the provider's limits apply. CrewKite sets no API key and never falls back to paid API billing. The provider section above has the dated line for each family, and Provider routes has the longer notes.
No. Connect each provider once on this computer with its Connect button in Settings. After that CrewKite reconnects on its own at every launch, and Local also reconnects by itself when LM Studio restarts. The OpenAI client runs in CrewKite's own home with its own sign-in, so nothing it does shows in your own OpenAI app.
Yes. A side question runs beside the builder on the model you pick, such as gpt-oss-20b in LM Studio, and never stops it.
Pick a model from another provider and send, or let the crew move the work when a provider runs out of included allowance. The next connected provider starts from your messages, the earlier answers, the list of changed files and the patches that still fit, in the same copy of your repository. Continuity and recovery has the details.
Only where your provider's own client sends it when you run a turn on that provider. CrewKite has no telemetry, no account and no server of its own, and nothing is hosted. A Local thread stays on your computer. Every thread works in its own copy, and your checkout changes only when you accept a branch.
OpenAI, Anthropic and Local run through their own clients on their supported paths. Google is different. Google's published FAQ says third-party access through its coding client login breaks its terms. CrewKite has no exception. Google stays off until you read a short summary in Settings and accept the risk. Use it only if that trade is yours to make. The dated links are in the docs.
The Linux preview runs on x64 and arm64 as DEB and RPM, with OpenAI, Anthropic and Local. Google is not on Linux. Windows is not supported yet. WSL2 is untested, so I cannot say whether the Linux preview runs there.
Choose Check for Updates in the app menu, or under Help on Linux. CrewKite also checks once a day. In 0.1.x an update is a download from the download page.
Quit the app, then remove the app, its data folder and its desktop profile. Thread refs in your repositories go with Delete in each thread's menu or one git command. The uninstall steps list every location.
The Apache License 2.0. The source, the docs and the architecture decision records are on GitHub.
Email me at ray@rootsnolimits.com. Open an issue on GitHub for bugs and requests. Report a security issue privately through GitHub as SECURITY.md describes, never by email or in a public issue.
macOS on Apple silicon, signed and notarized. Linux x64 and arm64 preview. Windows is not supported yet.