Index / Gitagent · updated Sep 3, 2026
Gitagent
open-gitagent/gitclaw · healthy · rank 28 of 43 by stars
A git-native agent framework where identity, rules, memory, tools, and skills are all version-controlled files — fork an agent, branch a personality, git log its memory. Recently hardened with security fixes and a Rust port (ira) powering a Tauri desktop app.
Facts
Repository
Runtime
Posture
Community
Security breakdown
Composite 78 / 100 · how these are scored
higher is safer
higher is safer
higher is safer
higher is safer
higher is riskier
Model access
Read from the repository, not written by a model · 57 files examined
3 providers · runs locally · custom endpoint
Pinned models
gpt-4o-mini released 2024-07-18 — 2 years old (from the public model catalogue)
Pin last edited 15 days ago
Evidence
Decision
Why choose Gitagent over OpenClaw?
Why choose this
- Agent state is fully version-controlled — memory, rules, and identity are diffable git files
- Recent security hardening closed load-time RCE and tool race conditions
- Rust port with embedded UI and Tauri desktop app for lower footprint
Tradeoffs
- Much smaller community and zero Reddit presence
- Plugin and skill ecosystem is young compared to OpenClaw's registry
- Shell/CLI tool access remains a meaningful risk surface
Best fit
- Developers who want agent config and memory as auditable git history
- Teams forking and diffing agent personalities like code
- Users wanting a local desktop agent app with model management
Avoid if
- You need a large, battle-tested community and plugin registry
- You want zero shell/CLI tool exposure by default
- You need proven multi-tenant or team collaboration features
Evidence is moderate: the README and detailed commit history show active, security-conscious development, but there is no Reddit discussion and only indirect web mentions, so community sentiment is largely inferred.
AI layer reviewed Aug 24, 2026 · how this is written
Star activity
668 stars today
Overview
Gitagent reimagines the AI agent as a git repository: agent.yaml holds model and runtime config, SOUL.md defines personality, RULES.md sets behavioral constraints, and memory/, tools/, skills/, and hooks/ are all version-controlled files. This makes agents forkable, branchable, and auditable — you can git log an agent's memory or diff its rules like any codebase. Distribution is via npm with a one-command installer that launches a voice-enabled web UI on localhost:3333.
Architecturally, the project is transitioning from its TypeScript/Node core to a Rust rewrite ('ira') that ships as a Tauri desktop app (Lyzr Edgespace) with an embedded single-file web UI, Ollama model management with auto-install, and an onboarding wizard. It supports MCP clients (stdio and HTTP/SSE), declarative YAML tools, lifecycle hooks, and an OpenAI-compatible API surface for edge agents.
Compared to OpenClaw, Gitagent's differentiator is its git-native state model and recent security hardening — commits show fixes for a load-time RCE via shell-interpolated git clone, tool concurrency races, and resource leaks. However, its community and plugin ecosystem are far smaller, telemetry/session tracing is built in, and CLI tool access means shell risk remains a consideration.