Autopilot

Autopilot builds your software project while you are away. You write what you want in plain words and start it once. You get working software, tested, with a record of every step. It is for one developer, or for a person with an idea and a developer friend.

Give it a plan tonight. Wake up to a working app, and proof that it works. Claude Code (the coding agent) writes the code; Python code decides when a task is complete.

Claude Code alone and with Autopilot

The full comparison has ten rows. Use Claude Code alone for a small change, a question, or an open design.

How it works

flowchart LR
    accTitle: How Autopilot builds a plan
    accDescr: Each task runs in a new Claude Code session. Python runs the checks. A pass merges to main, a fail retries on a stronger model, and a blocked task becomes a question while the other tasks continue. At the end, a completion audit adds tasks for gaps.
    plan["Plan"] --> task["Next task"]
    task --> session["New Claude Code<br/>session"]
    session --> gate{"Python runs<br/>the checks"}
    gate -->|pass| merge["Merge to main"]
    gate -->|fail| retry["Retry on a<br/>stronger model"]
    retry --> session
    retry -->|blocked| park["Question in NEEDS-YOU.md,<br/>build the other tasks"]
    park --> task
    merge --> task
    merge -->|plan done| audit{"Completion<br/>audit"}
    audit -->|gaps| task
    audit -->|complete| app["Working app"]

Follow the build

ap watch shows the build in the terminal; the status page shows the same data in your browser (sample data).

Real results

Two real projects built from a plan:

• SecretScan finds leaked secrets in a repository.
• envguard checks .env files against a schema and finds secrets in the git history.

Both runs used a subscription. The envguard audits found 8 real bugs; Autopilot fixed all. envguard is private; numbers come from the run log. A lean rerun of envguard used 39.5 percent fewer tokens per task.

Install

One line. It installs what is missing for your user, with no administrator prompt. Windows (PowerShell):

irm https://raw.githubusercontent.com/manishjnv/AutoPilot/main/install.ps1 | iex

macOS, Linux or WSL:

curl -fsSL https://raw.githubusercontent.com/manishjnv/AutoPilot/main/install.sh | bash

On Windows pick Windows or WSL, not both; WSL only for the sandbox. With uv already installed, uv tool install dev-autopilot is the same. Open a project folder and type ap. The first run logs you in to Claude (a paid plan, Pro or Max), then starts. See What ap does. After a release, run ap upgrade. For a repair or a second machine, run ap install. Checksums and each step: Install with one script.

Example output

Shortened real output:

task P02-T01 attempt 1 failed: PROBLEM: skip/focus marker added in tests/test_walker.py
session #10 task P02-T01 model=sonnet attempt=2
task P02-T01 done (sonnet, $0.39)
audit: completion audit: 92% complete, 8 corrective tasks (FIX001)

Documentation

Good to know

• Version 0.3.0, an early release. MIT license.
• One CLI agent per session. See limits and runners.