Harmonist: agent pack with enforced hooks
Harmonist
A set of specialist agents and hooks for Cursor that keeps the assistant from ending a turn without review and a memory note
- 2,252 ★ on the repository
- Last commit
- MIT
Broad access
Why this level
- Hooks run Python and bash scripts on every subagent dispatch, file edit and session stop
- The integration script writes AGENTS.md, memory and hooks into the project, and memory holds notes about your decisions
Install
- Works in
Manual install
git clone https://github.com/GammaLabTechnologies/harmonist.git && python3 harmonist/agents/scripts/integrate.py --pack harmonist --project .Run in the project root. The Cursor option: paste the text of integration-prompt.md into the agent.
Access and checks
How we reviewBroad accessThe tool can change data, run commands or handle secrets.
- Hooks run Python and bash scripts on every subagent dispatch, file edit and session stop
- The integration script writes AGENTS.md, memory and hooks into the project, and memory holds notes about your decisions
This is third-party code. Review the repository files before installing.
What it does
Harmonist drops into a project as a subfolder and turns the assistant into a team. The AGENTS.md file sets the orchestrator protocol, a catalogue of 193 agents in 16 categories tells it whom to hand work to, and memory lives in validated records. The main difference is hooks. The stop hook checks whether reviewers such as qa-verifier ran and whether the session-handoff.md file was updated. If not, it returns a message to the assistant and does not let the turn finish. Memory is written only through the memory.py script: it validates the schema, rejects duplicates and scans the text for about 30 kinds of secrets, from AWS keys to Telegram tokens. The pack contents are listed in the MANIFEST.sha256 file. Before copying into a project, the upgrade script checks the hashes and rejects a swapped agent file.
Who it is for. For teams that work in Cursor or connect the pack to Claude Code through an adapter and want to build a team of narrow agents with shared memory around the assistant.
Good fit when
- Make the assistant run a security review before touching auth code
- Carry context between sessions through validated memory with correlation IDs
- Pick narrow specialists for the project: Solidity audits, WeChat mini programs or Roblox Luau
- Run a long task where the next session has to see the open incidents of the previous one
- Set up orchestration without Docker or databases
Not a fit when
- You need one small script or a one-off edit: hooks and protocol slow things down more than they help
- The assistant cannot dispatch subagents: the protocol relies on their runs and session markers
- You need a ready server orchestration framework with queues and a dashboard: Harmonist is markdown, Python and bash scripts
Example request
Integrate Harmonist into this project, pick agents for Next.js and PostgreSQL and show which hooks will be enabled.Limitations
The main integration targets Cursor. Adapters exist for Claude Code, Copilot, Windsurf, Aider, Gemini CLI and others, but strict turn checking depends on which hooks a given editor supports. Python 3.9 or newer and Git are required, while bash is optional.
The pack folder stays in the project as a subfolder and must not be unpacked into the root. The integration script creates AGENTS.md, memory in .cursor/memory, hooks and the .cursor/pack-version.json version file, so review the changes in a diff before committing.
The agent catalogue is wide, and some roles target the Chinese market. Whether GitHub is reachable for installation from Russia depends on your network.
How to disable. Remove the pack hooks and entries from the .cursor folder and the AGENTS.md file, then the harmonist folder from the project. The .cursor/pack-version.json file shows what was installed.
FAQ
How do I add Harmonist to a Cursor project?
Clone the repository into the project root as a subfolder, open the project in agent mode and paste the contents of the harmonist/integration-prompt.md file. The assistant asks about the domain and roles and builds AGENTS.md. When it finishes, start a new chat.
Can I install Harmonist without Cursor?
Yes, from the command line: `python3 harmonist/agents/scripts/integrate.py --pack harmonist --project .`. On Windows the same script runs through py -3, with no WSL needed.
What happens if the assistant keeps failing the stop hook?
The hook sends the turn back at most three times. After that it records an incident, and the next session shows it at startup.
How is Harmonist different from ordinary agent frameworks?
Frameworks like LangChain, CrewAI and AutoGen give orchestration primitives and leave rule compliance to the prompt. In Harmonist editor hooks check the rules, the pack contents are verified by hash, and memory rejects entries containing secrets.
Do I need a database or Docker?
No. The pack is markdown, the Python standard library and bash, with no third-party dependencies.
Discussion
No questions or reviews yet.
Related
PluginsA multi-agent harness that verifies results by artifacts: stop hooks, an independent judge and an event log for Claude Code, Codex, Cursor and more
A skill plus hooks that keep a coding agent on task and cut speculative defenses, unread checksums and scope creep
A CLI that lets agents in different terminals message, watch and spawn each other
oah links 80-plus skills, five roles and /commit, /review, /test-tdd commands into Claude Code, Cursor, Codex and other agent folders
A portable .agent folder with memory, skills and protocols that plugs into many coding agents and keeps its knowledge when you switch tools
One hundred ready agent teams for Claude Code, each with specialists, an orchestrator skill and domain skills, in English and Korean
Comparison with similar tools
Entries that share the main search query, and how they differ.
| Entry | Official | Repository stars | Main difference |
|---|---|---|---|
| Harmonist: agent pack with enforced hooks (this entry) | no | 2,252 | |
| oh-my-claudecode | no | 40,377 | multi-agent orchestration for Claude Code with commands and autopilot, with no hook-based turn checks or hash verification |
| agency-agents-zh | no | 21,684 | a catalogue of expert-role prompts, with no hooks, memory or orchestrator protocol |
README in short
The starting point: a request in a prompt does not guarantee that rules are followed, so editor hooks check them. Next to the pack sit the GUIDE_EN.md guide, an agent map and integrations for ten editors.