Get autopilot installed and execute your first task end-to-end.
Prerequisites¶
- Python 3.11 or later
- Claude Code installed and authenticated
- A project directory you want to automate
Install¶
pip install claude-autopilotuv pip install claude-autopilotVerify:
autopilot --versionSet up authentication¶
You need either a Claude API key or a Claude Code subscription token:
export ANTHROPIC_API_KEY=your-key-hereclaude setup-token
export CLAUDE_CODE_OAUTH_TOKEN=<token from above>Option A: Let autopilot write the plan¶
The fastest path if you have an existing project:
cd my-project
# Build a shipping roadmap (analyses your project first)
autopilot roadmap .
# Generate a task manifest from the roadmap
autopilot plan .
# Review .dev/sprint.md, then execute
autopilot sprint .When autopilot plan runs, it writes .dev/sprint.md with a list of tasks. A judge agent evaluates the plan — if approved, approved: true is set automatically. Then autopilot sprint executes the tasks.
Option B: Write the plan yourself¶
Create .dev/sprint.md in your project root:
---
name: my-project
approved: true
max_budget_usd: 5.0
---
### [ ] add-ruff
Add ruff to the project. Install it as a dev dependency in `pyproject.toml`,
add a `[tool.ruff]` config section with `line-length = 100`, and run
`uv run ruff check src/` to verify.
**Done**: `uv run ruff check src/` exits 0.Then run:
autopilot sprint .What happens during execution¶
- Autopilot reads
.dev/sprint.mdand finds unchecked tasks - Each task spawns a fresh Claude Code session with the task description as context
- Claude Code does the work — writes code, runs commands, commits
- The worker verifies the task checkbox is checked in the manifest
- Failed tasks are retried up to
max_task_attemptstimes
Progress is printed to the terminal. The manifest is updated live — you can open it any time to see what's done.
Next steps¶
- Plan & Sprint scenario — the full plan-then-execute workflow
- Manifest Format — complete task syntax reference
- CLI Reference — all commands and flags