Quickstart

Prerequisites

Before you connect, make sure you have:

An Autonomy account with access to Studio.
At least one repository connected to Autonomy (the Autonomy GitHub App installed on it). New repos start in PENDING_INITIALIZATION - the setup step walks you through getting one to READY.
An MCP client - Claude Code or Claude Desktop.

The two MCP servers

You connect two MCP servers. They do different jobs and complement each other:

Server	What it is	What it does
`autonomy`	The product MCP	Does the work - projects, tasks, setup, Full Mode.
`autonomy-docs`	The docs MCP	Answers “how do I…” against this documentation, without leaving your editor.

Wiring both means your agent can read the docs and act on the platform in the same session.

Connect from Studio (recommended)

The fastest path is the “Connect your AI tools (MCP)” card in Studio’s User Settings. It gives a copyable Claude Code command, pointed at the hosted URL and wired to the OAuth login/consent flow.

Connect manually

Claude Code
Claude Desktop

# Product MCP (drives the platform - runs the OAuth login flow)
claude mcp add --transport http autonomy https://mcp.autonomyai.io/mcp

# Docs MCP (queries these docs)
claude mcp add --transport http autonomy-docs https://docs.autonomyai.io/_mcp/server

Add both servers to your client’s MCP config:

{
  "mcpServers": {
    "autonomy": {
      "url": "https://mcp.autonomyai.io/mcp"
    },
    "autonomy-docs": {
      "url": "https://docs.autonomyai.io/_mcp/server"
    }
  }
}

On first use of the autonomy server your client opens a browser to sign in and consent (your existing Autonomy identity and project permissions apply). The docs server is read-only and needs no auth.

Run your first task

Once connected, ask your agent to drive it:

Pick a project

Call list_projects and choose a project_id + workspace_id. If a just-connected repo isn’t READY, call initialize_project and poll list_projects until it is.

Start the task

start_task(project_id, workspace_id, instruction) returns a task_id immediately - the work runs in the background.

Watch it

Poll get_task(project_id, task_id) until the status is terminal or it needs input. Read the agent’s work with get_task_messages and the code changes with get_task_pr_diff.

Ship it

When you’ve approved the changes, complete_task opens or updates the pull request for your developers.

If you’re unsure how to drive a flow, the product MCP ships built-in guides: call load_autonomy_skill (names: setup, full-mode, tasks, troubleshooting) or use the /autonomy-getting-started prompt.

A full example, start to finish

Here’s a complete run - from a plain-English ask to an open pull request. You talk to your agent in natural language; under the hood it makes these calls:

You:   "In my storefront repo, add a dark-mode toggle to the settings page."

Agent: list_projects()
       → picks project_id=… , workspace_id=…  (status READY)

       start_task(project_id, workspace_id,
                  instruction="Add a dark-mode toggle to the settings page")
       → task_id=…  (status: running)

       get_task(project_id, task_id)   # polled until it settles
       → status: awaiting_input / completed

       get_task_pr_diff(project_id, task_id)
       → shows the changed files for you to review

You:   "Looks good - ship it."

Agent: complete_task(project_id, task_id)
       → opens the pull request for your developers ✅

With the skill installed, your agent runs this sequence correctly on its own - including waiting for your approval before complete_task.

Getting Started

Concepts

Tool Reference

Help

Prerequisites

The two MCP servers

Connect from Studio (recommended)

Connect manually

Run your first task

A full example, start to finish

Next steps

Projects & workspaces

Run a coding task

Full Mode

Troubleshooting

​Prerequisites

​The two MCP servers

​Connect from Studio (recommended)

​Connect manually

​Run your first task

​A full example, start to finish

​Next steps

Projects & workspaces

Run a coding task

Full Mode

Troubleshooting

Prerequisites

The two MCP servers

Connect from Studio (recommended)

Connect manually

Run your first task

A full example, start to finish

Next steps