What is Claude Project Autonomy?

It is an open-source CLI that turns the current repository into a persistent, low-prompt Claude Code workspace. Once enabled for a project, Claude can read and edit files, run commands, install dependencies, run tests and builds, and use the web without stopping for routine approvals.

Is autonomy configured per project or per machine?

Per project. The configuration is project-specific rather than machine-wide, so trust granted to one repository does not extend to the rest of your machine. On macOS, Linux, and WSL2 it also uses Claude Code's Bash sandbox to keep shell activity around the project workspace.

Yes. Run /autonomy off inside Claude Code. The tool tracks the configuration it added and removes only the values it owns. You can also run /autonomy status to inspect the current state or /autonomy doctor to diagnose unexpected permission prompts.

On native Windows /autonomy exits with code 3 instead of enabling. Is that a bug?

No, that is expected on native Windows. Because Windows has no Bash sandbox boundary, autonomy requires a one-time acknowledgement first. Run 'claude-autonomy enable --accept-no-sandbox' in your terminal before starting the Claude session, then invoke /autonomy normally. The acknowledgement is remembered per project. For stronger isolation, run the project in WSL2 or a container.

Claude Project Autonomy: Stop Approving Every Routine Claude Code Action

Claude Code is powerful enough to work through large development tasks, but long autonomous sessions can quickly turn into a sequence of permission prompts.

Run a command. Approve it. Edit a file. Approve it. Call an API. Approve it. Run the tests. Approve it again.

I built Claude Project Autonomy to make that workflow simpler.

The idea is straightforward:

I trust this project. Stop asking. Build.

Claude Project Autonomy turns the current repository into a persistent, low-prompt Claude Code workspace. Once enabled, Claude can perform normal development work without repeatedly asking for authorization.

What it is for

After enabling autonomy for a project, Claude can handle routine development operations such as:

reading and editing project files
creating and deleting files
running shell commands
installing dependencies
starting development servers
running tests, builds, linters, and scripts
using web search and web requests
debugging and iterating across multiple tool calls
spawning agents when needed

The configuration is project-specific rather than machine-wide.

On macOS, Linux, and WSL2, the tool also uses Claude Code's Bash sandbox so normal shell activity can remain constrained around the project workspace.

Install

Install the CLI globally:

terminal ~

npm install -g claude-project-autonomy

Then install the global Claude skill and helper configuration:

terminal ~

claude-autonomy install

You only need to perform this installation once.

Use it in any project

Open a repository normally:

terminal ~

cd my-project
claude

Then inside Claude Code run:

terminal ~

/autonomy

Autonomy is enabled for that project.

You can now give Claude a normal development task:

terminal ~

Investigate the failing API tests, fix the underlying issue,
run the full test suite, and verify the application.

Instead of interrupting the workflow for every routine action, Claude can continue working with the permissions already granted to the project.

Native Windows: one extra step

On macOS, Linux, and WSL2 the steps above are all you need.

On native Windows there is no Bash sandbox boundary, so the first time you run /autonomy the tool does not enable immediately. Instead it stops with exit code 3 and asks you to acknowledge that shell commands will run under your normal Windows user account:

terminal ~

Native Windows detected.
claude-autonomy enable --accept-no-sandbox

This is expected — not a crash. Run that command once in your terminal before starting the Claude session:

terminal ~

claude-autonomy enable --accept-no-sandbox

After the acknowledgement is recorded, start Claude and run /autonomy normally. It is remembered per project, so you only do this once.

For stronger isolation on Windows, run the project in WSL2 or a container instead.

Turn it off

Autonomy is reversible.

terminal ~

/autonomy off

The tool tracks the configuration it added and removes only the values it owns.

You can also inspect the current state:

terminal ~

/autonomy status

Or diagnose unexpected permission prompts:

terminal ~

/autonomy doctor

Why not just disable every permission check?

Claude Code supports more aggressive permission-bypass modes, but giving an agent unrestricted access to an entire machine is a different security model.

Claude Project Autonomy is designed around a narrower idea:

terminal ~

Current project       → trusted for development
Routine tool calls    → allowed
Rest of the machine   → outside the intended workspace boundary

It is not a promise that every Claude Code safeguard can be removed. Organization policies, explicit deny rules, protected paths, operating-system restrictions, and other platform safeguards can still take precedence.

Native Windows also does not currently provide the same built-in Bash sandbox boundary available on macOS, Linux, and WSL2. For stronger isolation on Windows, WSL2 or a container is recommended.

Open source

Claude Project Autonomy is open source and available on GitHub:

github.com/Audran-wol/Claude-autonomy

The project is still early, and feedback is welcome — especially from developers running Claude Code on longer autonomous tasks.

If you regularly find yourself clicking Allow while waiting for an agent to finish work you already asked it to do, this project is built for that workflow.