Skip to main content
GitAgent is an open-source framework for building AI agents that live in your git repository. Agents are defined as code, versioned with your project, and executed via a local runtime or CI pipeline — no separate deployment infrastructure required. GitAgent was originally called GitClaw. It has been renamed to reflect its core design: agents that are native to git workflows, not bolted on top of them.

​
What makes GitAgent different

Most agent frameworks treat git as a deployment mechanism — you write code, push it, and the framework picks it up. GitAgent inverts this: the agent’s definition is the repository. Skills, hooks, memory, and tools are all discovered by walking the repo tree according to the OpenGAP directory convention. This means:
  • No agent configuration UI — the repo is the source of truth
  • Agent diffs are code diffs — reviewing an agent change is reviewing a PR
  • Any CI system can run it — gitagent run works in GitHub Actions, GitLab CI, Jenkins, or locally

​
Core concepts

ConceptDescription
Agent manifestagent.yaml in the repo root — defines identity, model, and entrypoint
SkillsIndividual capabilities the agent can invoke (tools, knowledge, sub-agents)
HooksLifecycle callbacks — before/after each skill call, on error, on completion
PluginsExtend the runtime — add new skill types, output processors, auth providers
SkillsFlowCompose skills into multi-step workflows with branching and loops
MemoryPersistent memory that survives across runs (conversation history, learned facts)

​
How it works

repo/
├── agent.yaml              ← agent manifest
├── skills/
│   ├── search_web.py       ← a skill
│   ├── read_file.py        ← a skill
│   └── summarize.py        ← a skill
├── hooks/
│   └── on_complete.py      ← lifecycle hook
└── flows/
    └── research_flow.yaml  ← a SkillsFlow
When you run gitagent run, the runtime:
  1. Reads agent.yaml to get the agent’s identity and model
  2. Discovers all skills in skills/
  3. Loads hooks from hooks/
  4. Starts a conversation loop — the LLM decides which skills to call

​
When to use GitAgent

Good fit:
  • Agents that need to be version-controlled and code-reviewed
  • Teams that want agents as part of their existing git workflows
  • Agents that run on a schedule in CI (nightly analysis, weekly reports)
  • Self-hosted environments where cloud agent platforms aren’t an option
Consider alternatives:
  • If you need a visual no-code agent builder → Lyzr Studio
  • If you need multi-agent orchestration at scale → SuperFlow

​
Relationship to OpenGAP

GitAgent implements the OpenGAP specification. OpenGAP defines the directory structure, manifest format, and skill protocol that any compliant runtime must support. GitAgent is the reference implementation. This means agents you build with GitAgent are portable: any OpenGAP-compatible runtime can execute them.

​
Where to go next