GitAgent
An open-source, multi-model, git-native AI agent that lives inside your repository.
GitAgent is the most capable implementation of the OpenGAP standard. Your git repo is the single source of truth — session memory becomes branches, learning happens through a reinforcement learning loop, and every improvement is committed directly to your codebase. Written in TypeScript, MIT-licensed, self-hosted.
Install, run, watch it learn.
From install to first commit on a working branch in three commands.
How GitAgent works.
Core philosophy:
Clone and run
Clone a GitAgent-enabled repo or add GitAgent to an existing one. The agent definition lives in standard OpenGAP files: agent.yaml, SOUL.md, and a skills/ directory.
Session-as-branch
Every agent session creates a git branch. Working memory, decisions, and context are persisted as commits on that branch.
RL learning loop
GitAgent evaluates its own outputs, learns from feedback, and commits enhancements back — making the agent smarter over time without manual retraining.
Human-in-the-loop via PRs
Before any memory or skill update hits main, the agent opens a pull request for human review. Approve, reject, or modify just like you do with code.
Framework-agnostic execution
Built on the OpenGAP spec — run the same agent with Claude Code, OpenAI, Lyzr, CrewAI, or 8+ other adapters. Switch with a single CLI flag.
Production-grade by default.
Everything that makes GitAgent a production-grade AI coding agent runtime.
Git-native memory and state
Every session, decision, and learned behavior is a git commit. Full version history, diffs, blame, and rollback on your agent’s memory — not a black-box database.
Reinforcement learning loop
GitAgent doesn’t just execute — it learns. An RL loop evaluates performance, generates improvements, and commits them as skill or memory updates directly to the repo.
Multimodal capabilities
Handles text, images, and video inputs and outputs — going beyond text-only agents for richer, more capable workflows.
Built-in CLI tools
Ships with built-in tools including shell command execution, file operations, and more — ready to use out of the box.
Claude Code SDK-style API
Offers a familiar API interface modeled after the Claude Code Agent SDK, lowering the learning curve for developers in the Anthropic ecosystem.
Branch-based deployment
Promote agent changes through dev to staging to main, exactly like you ship software. Every environment gets its own agent version.
Agent forking and remixing
Fork a public agent, customize its SOUL.md, add new skills, and PR improvements back upstream — community-driven agent evolution.
Full audit trail
git diff shows what changed. git blame traces every line. Built-in compliance support for FINRA, SEC, and Federal requirements.
Multi-model support
Leverage different LLMs — Claude, OpenAI, Gemini, local models — for different tasks within the same agent definition.
Local and air-gapped deploy
Runs entirely on your infrastructure — no data leaves your network. Ideal for regulated industries where data sovereignty is non-negotiable.
Autonomous planning
Built-in task decomposition and planning for complex multi-step operations with full traceability.
Enterprise governance
Human-in-the-loop via PRs, kill switches, audit logging, and compliance frameworks for BFSI deployments.
What sets GitAgent apart.
Why teams choose GitAgent for production agent deployments.
Total transparency
Every thought, decision, and skill update is a git commit — diffable, reviewable, and reversible. No black-box memory, no opaque state.
Self-improving agent
GitAgent’s RL loop continuously improves the agent and commits those improvements for human review — your agent gets smarter without manual retraining.
No vendor lock-in
Built on the open OpenGAP standard — switch between Claude Code, OpenAI, Lyzr, CrewAI, or any of 8+ adapters without changing your agent definition.
Enterprise-grade governance
Built-in compliance frameworks for FINRA, SEC, and Federal requirements, with audit logging, human-in-the-loop via PRs, and kill switches.
True local and air-gapped deployment
Runs entirely on your infrastructure — no data leaves your network. Ideal for regulated industries where data sovereignty is non-negotiable.
Where GitAgent shines.
The deployments where git-native, self-improving agents deliver outsized value.
Enterprise AI assistants
Deploy compliant, auditable agents in banking, insurance, or healthcare where every decision must be traceable.
Developer copilots
Agents that learn your codebase, improve over time, and submit PRs with their enhancements.
Multi-model workflows
Leverage different LLMs (Claude, OpenAI, local models) for different tasks within the same agent definition.
Team-shared agents
Fork, customize, and collaborate on agents using the same git workflows your team already knows.
Regulated industries (BFSI)
Built-in compliance, audit trails, and kill switches for financial services deployments.
OpenGAP — the Open Git Agent Protocol.
GitAgent is built on OpenGAP, the open standard for git-native AI agents. Explore the protocol that makes your repository the single source of truth for any AI agent.
Start using GitAgent.
Install globally, point at your project, and let the agent work. Your repo is your agent.