ClawAI Coding Agent for VS Code

Every AI, one workspace — right inside VS Code.

ClawAI Coding Agent puts a chat and a coding agent in your editor. Ask a question, fix a bug, write tests, or hand over a bigger task. It reads your project, proposes changes, and waits for your OK before it touches a file.

It connects to your ClawAI account, so you can use many AI models from one place: hosted models, or models running on your own machine.

Version 1.100.0 delivers a more capable agent that can plan, edit files, run commands, use git, and drive a browser, all with your approval, and it keeps working with the classic chat-and-review flow if a newer feature is not available on your ClawAI server. The approval choices now mean what they say (Strict asks every time, Plan can still read your workspace), the chat box remembers Run, Context and Web research, and in Plan mode the agent answers that it cannot write instead of stopping with an error. Compare and Compare + Judge now show one card per model as each model starts and fill it while it writes, then show the judge's verdict with a winner, ranks and scores the moment it is ready. Scheduled tasks can now follow a calendar (for example weekdays at 9:00), and commands sent to a runner machine run with your sandbox setting and without your editor's private environment. The command-line agent can now be set up like the chat box (effort, what it is shown first, which web tools it may use, and how much it may do without asking), and the agent can read several pages of one website.

Why use it

• You stay in control. You see every change as a diff before it is applied, and you can undo it.
• Pick your model, or let ClawAI pick. Use automatic routing, or choose a cloud or local model yourself.
• Safe by default. Secret files are never read, commands are limited, and approvals are on until you say otherwise.
• Works with your project. It understands your files, selection, git state, and your own project rules.
• Private. Sign-in happens in your browser. A zero-retention mode keeps run history in memory and asks the server not to keep your content.
• Your language. The interface comes in 13 languages, including right-to-left Arabic and Persian.

What you can do

Chat and agent

• Chat in the ClawAI panel, in an editor tab, or with @clawai in VS Code Chat.
• Keep several chats open at once, and run two at the same time with different models.
• Choose Auto (the agent acts, with your approvals) or Plan mode (read-only: it thinks and suggests, but changes nothing).
• Compare two to five models on the same question, then ask a judge model to pick the best answer.
• Ask about a selection, the current file, or your whole workspace. Type src/app.ts:10-40 in a message to pull in exactly those lines.
• Ask a quick side question without derailing the main task.
• Choose how answers are written: default, concise, explanatory, or learning.

Files and code changes

• Generate code, fix a selection, review code, write tests, write docs, make a plan, or audit the whole workspace.
• Review changes as a normal VS Code diff before anything is written.
• Undo your last ClawAI edits, step by step.
• Create checkpoints and restore them later, or rewind a conversation to an earlier point.

Running commands safely

• The agent can run development commands (tests, builds, linters) in a visible VS Code terminal that you can cancel.
• Commands are checked first. Dangerous or unknown ones are refused or need your approval.
• Optional sandbox: run commands inside your operating system's sandbox or a Docker container, with network access off unless you turn it on.
• The agent can open pages in a browser to check your web app, and you can attach what it saw to your chat.

Approvals and permission modes

Pick the Approval level under More settings in the chat box:

• Plan: read-only. Anything that changes something is refused.
• Ask for Approval: the default. You approve each change and command.
• Auto Edit: routine file edits go through. Commands and riskier actions still ask.
• Autonomous Scoped: routine edits and commands inside the workspace go through. Committing, pushing, deleting and MCP calls still ask. Writing outside the workspace is refused in every level, never offered for approval.
• Strict: asks like Ask for Approval, and refuses deleting, production and admin-level actions.

No level skips Workspace Trust or secret-file protection, and you can always cancel. On an approval card, Always allow in this workspace remembers your choice for that trusted workspace.

Sessions, history and resume

• Your conversations are saved. Reopen, rename, archive, group, or restore them, and open one in a new window.
• Search your run history, export a transcript, and see a session recap.
• Resume a conversation that started somewhere else, such as the ClawAI web app.
• Long chats can be summarized (compacted) so they can keep going.
• See how much you have used with ClawAI: Show Usage.

Attachments: images, PDF and voice

• Paste, drop, or pick screenshots, images, videos, documents, PDFs, and source files into the message box.
• Attach terminal output with one command.
• Dictate your message by voice in the composer. If your setup does not allow it, ClawAI tells you how to use your system's dictation instead.
• Files whose names look like secrets (such as .env or id_rsa) are refused.

Git and pull requests

• The agent can look at your git changes, make commits, and help prepare pull requests when you allow git.
• Post a code review comment to GitHub or GitLab straight from VS Code.
• Optional Slack notifications about your runs.

Plugins, MCP and skills

• Add tools with MCP servers (local or remote) in your settings.
• Browse, install, enable, and disable plugins from plugin marketplaces.
• Teach ClawAI your project with skills, rules, and named sub-agents kept in a .clawai folder. Create it with ClawAI: Initialize .clawai.
• Run your own commands at set points in a run with hooks.

Workflows and scheduled tasks

• Run saved workflows from the Command Palette.
• Manage scheduled tasks and cloud routines that run on their own.
• Run a remote job on demand.

Remote and command-line use

• Start a cloud coding session, or register your machine as a runner.
• Pair the editor with your phone from a QR code, then run jobs you created for the paired device.
• Let ClawAI send commands to this editor after you confirm each one.
• A command-line runner (clawai -p "your task") is included in the extension (dist/headless.mjs) for scripts and CI. It needs Node.js 22.13 or newer and a ClawAI credential in environment variables.
• Some of these need a ClawAI backend that supports them, or a second machine. See Remote and automation.

Let the agent test and ship for you

The command-line agent (clawai -p "your task", also available as a code library) can do more than edit files. It can run your tests, start your app, click through it in a real browser, call your API, split a big job between helpers, then commit and push. These are command-line options, not editor settings, and every one of them is off until you turn it on.

By default the agent can only read your project and look at git. Turn on what the job needs, and nothing more:

• Run programs and tests: --allow-tools read,write,command
• Commit and push: add git-write
• Check your website in a real browser: --browser-allow-host localhost:3000
• Check your API: --http-allow-host localhost:3000 (add http-write to let it send changes)
• Keep a server running while it works: comes with command
• Look at a screenshot: --vision
• Follow a plan it cannot skip: --plan-file plan.json --require-plan
• Let helpers work in parallel: --allow-tools agents (only for big jobs with separate parts)
• Run shell scripts: --allow-tools shell --allow-shell --permission-mode ask (off unless you give all three)

Safety, in plain words:

• The shell is off. Nothing can run a free-form script unless you switch it on and approve each script.
• The browser and the API tool can only reach websites you name. Local and private addresses are blocked unless you list them.
• Your approval level decides what is asked. With no terminal to ask on (for example in CI), anything that needs approval is refused, never assumed.
• You can fence where it may write (--write-scope), and define what "done" means with checks that run real commands (--done-check) instead of trusting the agent's word.
• Passwords, tokens and keys are hidden from the agent and from its logs.

Every tool, with its limits and what is asked in each approval level, is in docs/TOOLS.md. Common questions are in docs/FAQ-AGENT-TOOLS.md; all options are in docs/HEADLESS.md.

Privacy and zero-retention

• You sign in through your browser. The extension never sees your password.
• Tokens are kept in VS Code's secure storage, never in settings or files.
• Turn on Zero data retention to keep run history in memory only, ask the server not to keep your content, and block uploads and sharing.
• Nothing is sent to a telemetry service unless you set one up yourself.

Requirements

• VS Code 1.98 or newer.
• A ClawAI account and a ClawAI server you can reach: Local (https://claw.local) or Cloud (https://claw-ai.co), or your own.
• A trusted workspace to let ClawAI change files. In an untrusted workspace, chat and read-only review still work.
• Optional: Docker, if you want commands to run in a container.

Quick start

1. Install ClawAI Coding Agent from the VS Code Marketplace.
2. Open the claw icon in the Activity Bar, or press Ctrl+Shift+A (Cmd+Shift+A on macOS).
3. Pick a server on the connection screen: Local, Cloud, or Custom.
4. Choose Connect to ClawAI. Your browser opens. Sign in and approve VS Code. You never type a password into the extension.
5. Ask something. Try "explain this file", or select some code and press Ctrl+Shift+Enter. Keep Automatic routing on, or pick a model.

Commands and shortcuts

Open the Command Palette (Ctrl+Shift+P) and type ClawAI.

Keyboard shortcuts marked "(selection)" work while text is selected in the editor. Review Selected Code and Generate Tests work while an editor has focus. You can also right-click a selection to ask, fix, or review it.

Settings that matter

Open Settings and search for clawAI.

Secrets such as tokens and API keys are never settings.

FAQ and troubleshooting

I can't connect. Check the server address on the connection screen. Local is https://claw.local and Cloud is https://claw-ai.co. Non-local servers must use https://. Run ClawAI: Show Logs for details.

The browser opens but VS Code stays signed out. Approve the request in the browser tab, then return to VS Code. If it still fails, run ClawAI: Log Out, then connect again.

ClawAI will not edit my files. Trust the workspace (VS Code shows a banner or use Workspaces: Manage Workspace Trust). In an untrusted workspace ClawAI only chats and reviews.

A command was refused. Commands are checked for safety. Approve it when asked, or check that you are in the right permission mode. Secret files and paths outside your workspace are always blocked.

I don't see any models. Run ClawAI: Refresh Models. Your account decides which models you can use.

How do I undo what the agent did? Run ClawAI: Undo Last ClawAI Edit (Ctrl+Alt+Z), or restore a checkpoint.

See it in action

These pictures use made-up sample data. Each one also comes in a light version in the same folder.

Help and links

• Report a problem or ask a question: use ClawAI: Send Feedback, or open an issue on GitHub.
• What changed: CHANGELOG.
• Security and privacy: SECURITY.md.
• Project rules and skills: the .clawai folder guide.
• Technical details for developers: ARCHITECTURE.

License

MIT