Before agent-fox

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.

With agent-fox

You write the same spec, run af 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 commits to develop.

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

Quick Start

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

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

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

# Create the task graph from your specs
af plan

# Run autonomous coding sessions
af code 

# Check results
af standup

See the CLI reference for all command options.

Night Shift — Fix-Only Daemon

Keep your codebase healthy while you sleep. Night Shift is a continuously-running fix-only daemon that polls for af:fix-labelled GitHub issues and autonomously processes them through a three-stage pipeline (triage → coder → reviewer).

# Start the fix daemon (Ctrl-C to stop gracefully)
af night-shift

Installation

uv tool install af --from git+https://github.com/agent-fox-dev/agent-fox.git#subdirectory=packages/af

Development

The repository is a uv workspace with five packages:

Package	Description
packages/af/	CLI for the agent-fox orchestrator (af command)
packages/agentfox/	Core library — spec engine, graph planner, session runtime, workspace tools
packages/afspec/	Standalone library for the agent-fox specification format (v1.2)
packages/agentspec/	AI-powered spec creation library
packages/spec/	CLI for AI-powered spec creation (spec command)

af  ──▶  agentfox  ──▶  afspec
              ▲
spec ──▶ agentspec ──┘──▶  afspec

uv sync                      # install all packages in editable mode

Command	What it does
make check	Lint + all tests (use before committing)
make test	All tests
make test-unit	Unit tests only
make test-property	Property-based tests only
make test-integration	Integration tests only
make lint	Check lint + formatting
make format	Auto-format code

Changes are immediately reflected via editable install. To run the local version explicitly (rather than a globally installed release):

uv run af <command>

Documentation

Full documentation lives in docs/:

• CLI Reference — all commands, flags, and exit codes
• Configuration Reference — every config.toml option (all sections and fields)
• Agent Archetypes — archetype registry, modes, convergence
• Skills — bundled Claude Code slash commands (/af-spec)

For a deeper understanding of the system's internals — how specs become task graphs, how agents are dispatched in parallel, how the knowledge store works, and how night-shift processes fix issues — see the Architecture Guide.

References

agent-fox draws on ideas from the following research:

• MAGMA — A multi-graph memory architecture for AI agents. agent-fox's knowledge system uses a similar approach: typed facts with causal links, embedding-based retrieval, and lifecycle management (deduplication, contradiction detection, decay). arXiv:2601.03236

• Sleep-time Compute — Explores how pre-computation outside of inference time can improve agent performance. Night-shift's autonomous maintenance model applies this principle: the system does useful work while the developer is away, so the codebase is healthier when they return. arXiv:2504.13171

• Memory in the Age of AI Agents: A Survey — A comprehensive survey of memory architectures for AI agents. Provides context for agent-fox's design choices around fact extraction, supersession, and retrieval. GitHub

---

Built exclusively for Claude Code. And mostly by agent-fox.