A ready-to-use library of AI development agents that work across Claude, Codex, OpenCode, Cursor, and GitHub Copilot. Install once and you get a full software development workflow — planning, implementation, review, testing, auditing, and docs — driven by a handful of agents you talk to directly.
Install the agents into your own harness with one command from the repository root:
python3 deploy_agents.pyThe first run asks which harnesses you use (Claude, Codex, OpenCode, Cursor, GitHub) and remembers your choice. It only writes files this system generated — your hand-maintained config is never touched.
→ See INSTALLATION.md for full installation instructions, options, and destinations.
Put the agent name in the prompt. The @name form is a repository routing
convention, so it works after these assets are deployed:
codex '@feature-decomposer decompose Phase 08a into execution-ready feature bundles'You can also say Act as the feature-decomposer ... in plain language. Do not
use codex -p feature-decomposer: -p/--profile selects a Codex configuration
profile and does not select a custom agent.
You interact with a small set of primary agents. Each one can drive a fleet of automated helpers behind the scenes, so you stay focused on a few entry points.
A numbered pipeline you step through for building a project end to end:
| Step | Agent | You do |
|---|---|---|
| 1 | Project - Planner | Describe your project → get a phased roadmap |
| 2 | Phase - Refiner | Refine and stress-test one phase |
| 3 | Feature - Decomposer | Break the phase into ready-to-build features |
| 4 | Phase - Execute | Kick off — implementation, review, and QA run hands-free |
| 5 | PR - Review | Get a plain-language readiness report on your diff |
You drive these steps in order.
Reach for these individually, whenever you need them — no pipeline required:
- Single Feature - Agent — a small, scoped change with an approval gate before it edits
- Debugger — diagnose and fix a frontend or backend error
- Docs Writer — create or update your repo's documentation
- Web Researcher — research a topic and produce a cited findings report
- Audit - Code, Infra, Refactor — health-check your code, infra, or structure
- Test - Orchestrator — analyze, write, or fix your test suite
- Prod Code Review — a final GO / NO-GO readiness gate
- Security Scan — a full-codebase security assessment
- Unity Reviewer / Visual Verifier — review Unity C# and verify what actually renders
- Eval - Grader / Instructions Manager — score agent runs and manage AI instruction files
Behind these, a set of automated subagents and on-demand skills do the detailed work —
you never invoke them directly. The library ships 42 source agent definitions in
source_of_truth/agents/. For the complete catalog and how the pipeline flows,
see source_of_truth/agents/README.md.
For users there are two steps, and the tool handles both:
- Pick your harnesses — the first run of
deploy_agents.pyasks which tools you use and saves the choice. - Deploy — it copies the agents into the real config directories each harness reads
(
~/.claude,~/.codex,~/.config/opencode,~/.cursor, and this repo's.github/for Copilot).
Deploy is safe by construction: a destination file is only ever overwritten or removed
when it positively carries a generated marker (or lives inside a generated skill
directory). Hand-maintained files are never touched. Re-run deploy_agents.py anytime
to pull the latest.
Deploy also maintains a baseline instructions file per harness — CLAUDE.md for
Claude, AGENTS.md for Codex and OpenCode, an always-applied rule for Cursor, and
.github/copilot-instructions.md for Copilot. It carries three managed sections
(Context7 usage, code-review-graph usage, and agent/skill discovery) wrapped in HTML
sentinel comments; deploy splices only those sections, so anything you write outside
them in the same file is preserved.
python3(standard library only — no third-party runtime dependencies)- Optional harness tooling depending on what you deploy to: Claude Code, Codex, OpenCode, Cursor, or GitHub Copilot
- VS Code if you want the Copilot agent picker
The deploy script also installs and configures two optional companion tools the agents
use when present: code-review-graph
(via pip/pipx) and the Context7 MCP server (via npx,
requires Node.js). Pass --skip-tools to opt out; a failed tool install never blocks
asset deployment.
python3 deploy_agents.py # use saved selection, or prompt (tty) and save
python3 deploy_agents.py --harness claude,cursor
python3 deploy_agents.py --all
python3 deploy_agents.py --list # show harnesses and resolved destinationsThe first interactive run saves your choice to .deploy-config.json (gitignored);
subsequent runs are just python3 deploy_agents.py. Full details are in
INSTALLATION.md.
- INSTALLATION.md — how to deploy the agents into your harness
- docs/COPILOT_SETUP.md — using the agents with GitHub Copilot (open this repo in your VS Code workspace alongside your project)
- source_of_truth/agents/README.md — the full agent catalog and pipeline flow
- CONTRIBUTING.md — for maintainers and contributors: how the repo is authored, transformed, and laid out
- code-review-graph by tirth8205 — a local-first code knowledge graph (MCP + CLI) the agents use for token-efficient, structure-aware code review. The deploy script installs and configures it automatically.
- Context7 by Upstash — an MCP server providing current, version-accurate library documentation. Also auto-configured by the deploy script.