It reads your repo, not a description of it
Connect GitHub from Settings and pick a repo. It gets cloned into a sandboxed workspace, so the agent answers questions about the code you actually have, including the parts you forgot were there.
ClawBot Crew · Coding Agent
The Coding Agent works inside a GitHub repo you own. It reads your actual files, asks what it does not understand, writes a short plan, then builds in small steps and opens a pull request. You review that PR exactly like you would a colleague's — because until you merge it, nothing about your project has changed.
Your repo, your branch protection, your merge button. Disconnect GitHub in one click and the code it already wrote stays yours.
It behaves like a contributor who read the codebase first, not a generator that produces files in a vacuum.
Connect GitHub from Settings and pick a repo. It gets cloned into a sandboxed workspace, so the agent answers questions about the code you actually have, including the parts you forgot were there.
It clarifies what you meant, writes a short plan you can read in under a minute, and then builds in small steps you can follow. If the plan is wrong, you find out before the code exists rather than after.
Work arrives as a PR against your repo — a diff, a description, and the test run behind it. Your branch protection rules and reviewers apply exactly as they always did.
Tokens and secrets stay locked inside your own private workspace. The agent uses them, and cannot print them back to you or to anyone else in a conversation.
From a sentence to a reviewable diff.
In plain English. "Add 2FA." "Our webhook handler drops retries." "Move the auth middleware into its own package." You do not need to know which files are involved.
YouIt goes through the relevant parts of the repo and comes back with the questions it genuinely cannot answer from the code. Answering two questions here is what keeps it from building the wrong thing for an hour.
AgentThe plan is short and concrete: what changes, in what order, and what it is deliberately not touching.
Then it implements step by step, running your tests as it goes. When a test fails it fixes the code rather than deleting the test.
AgentYou get a normal pull request. Read the diff, ask for changes in the conversation, merge when you are happy. Nothing reaches your main branch on its own.
YouAnything from a one-line fix to a feature with a spec attached.
Worth reading before you point it at something that matters.
Every change arrives as a pull request. That is the whole safety model — if you approve PRs without reading them, this agent is exactly as dangerous as any other contributor you do that to.
It runs the tests you have. On a repo with thin coverage, a green run means less than it looks like, and the review burden lands back on you.
It is at its best on well-scoped changes — a feature, a refactor, a bug with a reproduction. "Rewrite the backend" produces a plan you should argue with, not a weekend of unattended work.
The free trial connects a single repo. Unlimited repos and parallel task runs come with the paid subscription.
One subscription per crew agent, with its own monthly credit allowance.
Includes 1,500 credits per month, unlimited connected repos, and parallel task runs. Free trial available, limited to one repo.
Crew agents are hired on top of a paid ClawBot plan. One base plan covers every crew agent you hire.
No. Work comes back as a pull request against your repo, and your existing branch protection and review rules apply unchanged.
Yes, with one caveat worth being honest about. You can describe what you want and get a working app shipped to your GitHub without writing code.
But reviewing a pull request is the safety mechanism, and you will be approving diffs you cannot fully read. Start with small, low-stakes changes so you can judge the results by whether the app does what you asked.
Only the repos you connect. Each one is cloned into a workspace scoped to your account.
A top-tier coding model, chosen for this work rather than for being cheap. That is also why the agent has its own credit allowance instead of running unbounded.
Each crew agent is its own subscription, with its own memory and its own credits. They can hand work to each other when you tag them in chat.
Not a toy task. Give it something from your actual backlog with a reproduction, and read the pull request it opens. That is the honest test.