Point agent-fox at a spec. Walk away. Come back to working code across 50+ commits.

agent-fox is an autonomous coding-agent orchestrator built exclusively for Claude. It reads your specifications, builds a dependency graph of tasks, and drives Claude coding agents through each one — in parallel, in isolated worktrees, with structured memory, adaptive model routing, and multi-archetype review pipelines.

You write a spec, then sit in front of your terminal babysitting an AI agent for hours. You paste context, fix merge conflicts, restart after crashes, and lose track of what's done.

By session 10 you're exhausted and the agent has forgotten everything from session 1.

You write the same spec, run agent-fox code, and go do something else.

The fox reads your specs, plans the work, spins up isolated worktrees, runs each session with the right context, handles merge conflicts, retries failures, extracts learnings into structured memory, and merges clean commits to develop.

You come back to a finished feature branch and a standup report.

uv tool install agent-fox

Or install directly from the repository:

uv tool install git+https://github.com/agent-fox-dev/agent-fox.git

# Initialize your project (use --skills to install Claude Code skills)
agent-fox init --skills

# Create the task graph from your specs
agent-fox plan

# Run autonomous coding sessions with 4 agents in parallel
agent-fox code --parallel 4

# Check progress
agent-fox status

See the CLI reference for all command options.

Keep your codebase healthy while you sleep. Night Shift is a continuously-running maintenance daemon that hunts for linter debt, dead code, test coverage gaps, outdated dependencies, and more — then files GitHub issues and autonomously fixes the ones labelled af:fix.

# Start the maintenance daemon (Ctrl-C to stop gracefully)
agent-fox night-shift

# Automatically label every discovered issue as af:fix for hands-off repair
agent-fox night-shift --auto

Requires a GitHub platform configured in .agent-fox/config.toml ([platform] type = "github"). See the Night Shift CLI reference and configuration for full details.

Your project needs specs under .specs/ before running plan or code.

Use the /af-spec skill in Claude Code to generate them from a PRD, a GitHub issue or a plain-English description:

/af-spec [path-to-prd-or-prompt-or-github-issue-url]

agent-fox ships with a set of Claude Code skills that assist with spec authoring, architecture decisions, code simplification, and more.

Full documentation lives in docs/:

• CLI Reference — all commands, flags, and exit codes
• Configuration Reference — every config.toml option (all sections and fields)
• Archetypes — agent roles (Coder, Skeptic, Oracle, Auditor, Verifier, …)
• Skills — bundled Claude Code slash commands (/af-spec, /af-fix, …)

The Scope Guard subsystem prevents wasted coder sessions by enforcing that test-writing task groups produce only stubs (not full implementations) for non-test code.

When a task group has a test-writing archetype, the system:

1. Prompt Injection — Injects a SCOPE_GUARD:STUB_ONLY directive into the coder session prompt, instructing the agent to produce only type signatures and stub bodies for all non-test code.

2. Post-Session Validation — After the session completes, scans all non-test source files modified by the session and verifies that every new or modified function body contains only stub placeholders.

3. Violation Reporting — If non-stub implementations are found in non-test code, the session is flagged with a stub-enforcement violation and a warning is emitted to the operator.

Stub detection supports the following languages and placeholder patterns:

A function body is classified as a stub only if its entire content (after stripping comments and whitespace) consists of a single recognized placeholder. Any additional statements alongside a placeholder disqualify it as a stub.

Stub enforcement applies only to production code. Functions inside language-appropriate test blocks are excluded:

• Rust — code inside #[cfg(test)] modules or #[test] functions
• Python — files matching test_*.py / *_test.py, or functions starting with test_, or classes inheriting from unittest.TestCase
• TypeScript/JavaScript — files matching *.test.ts / *.spec.ts / *.test.js / *.spec.js, or code inside describe() / it() / test() blocks

Files in unsupported languages are skipped with a warning logged.

uv sync --group dev
make test              # all tests
make lint              # check lint + formatting
make check             # lint + all tests

uv sync installs the project in editable mode, so changes you make to the source are immediately reflected when you run agent-fox. To run the local version explicitly (rather than a globally installed release):

uv run agent-fox <command>