gitclaw

Gitclaw

A universal git-native multimodal always learning AI Agent (TinyHuman)
Your agent lives inside a git repo — identity, rules, memory, tools, and skills are all version-controlled files.

Install • Quick Start • SDK • Architecture • Tools • Hooks • Skills

Why Gitclaw?

Most agent frameworks treat configuration as code scattered across your application. Gitclaw flips this — your agent IS a git repository:

• agent.yaml — model, tools, runtime config
• SOUL.md — personality and identity
• RULES.md — behavioral constraints
• memory/ — git-committed memory with full history
• tools/ — declarative YAML tool definitions
• skills/ — composable skill modules
• hooks/ — lifecycle hooks (script or programmatic)

Fork an agent. Branch a personality. git log your agent's memory. Diff its rules. This is agents as repos.

One-Command Install

Copy, paste, run. That's it — no cloning, no manual setup. The installer handles everything:

bash <(curl -fsSL "https://raw.githubusercontent.com/open-gitagent/gitclaw/main/install.sh?$(date +%s)")

This will:

• Install gitclaw globally via npm
• Walk you through API key setup (Quick or Advanced mode)
• Launch the voice UI in your browser at http://localhost:3333

Requirements: Node.js 18+, npm, git

Or install manually:

npm install -g gitclaw

Quick Start

Run your first agent in one line:

export OPENAI_API_KEY="sk-..."
gitclaw --dir ~/my-project "Explain this project and suggest improvements"

That's it. Gitclaw auto-scaffolds everything on first run — agent.yaml, SOUL.md, memory/ — and drops you into the agent.

Local Repo Mode

Clone a GitHub repo, run an agent on it, auto-commit and push to a session branch:

gitclaw --repo https://github.com/org/repo --pat ghp_xxx "Fix the login bug"

Resume an existing session:

gitclaw --repo https://github.com/org/repo --pat ghp_xxx --session gitclaw/session-a1b2c3d4 "Continue"

Token can come from env instead of --pat:

export GITHUB_TOKEN=ghp_xxx
gitclaw --repo https://github.com/org/repo "Add unit tests"

CLI Options

Flag	Short	Description
--dir <path>	-d	Agent directory (default: cwd)
--repo <url>	-r	GitHub repo URL to clone and work on
--pat <token>		GitHub PAT (or set GITHUB_TOKEN / GIT_TOKEN)
--session <branch>		Resume an existing session branch
--model <provider:model>	-m	Override model (e.g. anthropic:claude-sonnet-4-5-20250929)
--sandbox	-s	Run in sandbox VM
--prompt <text>	-p	Single-shot prompt (skip REPL)
--env <name>	-e	Environment config

SDK

npm install gitclaw

import { query } from "gitclaw";

// Simple query
for await (const msg of query({
  prompt: "List all TypeScript files and summarize them",
  dir: "./my-agent",
  model: "openai:gpt-4o-mini",
})) {
  if (msg.type === "delta") process.stdout.write(msg.content);
  if (msg.type === "assistant") console.log("\n\nDone.");
}

// Local repo mode via SDK
for await (const msg of query({
  prompt: "Fix the login bug",
  model: "openai:gpt-4o-mini",
  repo: {
    url: "https://github.com/org/repo",
    token: process.env.GITHUB_TOKEN!,
  },
})) {
  if (msg.type === "delta") process.stdout.write(msg.content);
}

SDK

The SDK provides a programmatic interface to Gitclaw agents. It mirrors the Claude Agent SDK pattern but runs in-process — no subprocesses, no IPC.

query(options): Query

Returns an AsyncGenerator<GCMessage> that streams agent events.

import { query } from "gitclaw";

for await (const msg of query({
  prompt: "Refactor the auth module",
  dir: "/path/to/agent",
  model: "anthropic:claude-sonnet-4-5-20250929",
})) {
  switch (msg.type) {
    case "delta":       // streaming text chunk
      process.stdout.write(msg.content);
      break;
    case "assistant":   // complete response
      console.log(`\nTokens: ${msg.usage?.totalTokens}`);
      break;
    case "tool_use":    // tool invocation
      console.log(`Tool: ${msg.toolName}(${JSON.stringify(msg.args)})`);
      break;
    case "tool_result": // tool output
      console.log(`Result: ${msg.content}`);
      break;
    case "system":      // lifecycle events & errors
      console.log(`[${msg.subtype}] ${msg.content}`);
      break;
  }
}

tool(name, description, schema, handler): GCToolDefinition

Define custom tools the agent can call:

import { query, tool } from "gitclaw";

const search = tool(
  "search_docs",
  "Search the documentation",
  {
    properties: {
      query: { type: "string", description: "Search query" },
      limit: { type: "number", description: "Max results" },
    },
    required: ["query"],
  },
  async (args) => {
    const results = await mySearchEngine(args.query, args.limit ?? 10);
    return { text: JSON.stringify(results), details: { count: results.length } };
  },
);

for await (const msg of query({
  prompt: "Find docs about authentication",
  tools: [search],
})) {
  // agent can now call search_docs
}

Hooks

Programmatic lifecycle hooks for gating, logging, and control:

for await (const msg of query({
  prompt: "Deploy the service",
  hooks: {
    preToolUse: async (ctx) => {
      // Block dangerous operations
      if (ctx.toolName === "cli" && ctx.args.command?.includes("rm -rf"))
        return { action: "block", reason: "Destructive command blocked" };

      // Modify arguments
      if (ctx.toolName === "write" && !ctx.args.path.startsWith("/safe/"))
        return { action: "modify", args: { ...ctx.args, path: `/safe/${ctx.args.path}` } };

      return { action: "allow" };
    },
    onError: async (ctx) => {
      console.error(`Agent error: ${ctx.error}`);
    },
  },
})) {
  // ...
}

QueryOptions Reference

Option	Type	Description
prompt	string | AsyncIterable	User prompt or multi-turn stream
dir	string	Agent directory (default: cwd)
model	string	"provider:model-id"
env	string	Environment config (config/<env>.yaml)
systemPrompt	string	Override discovered system prompt
systemPromptSuffix	string	Append to discovered system prompt
tools	GCToolDefinition[]	Additional tools
replaceBuiltinTools	boolean	Skip cli/read/write/memory
allowedTools	string[]	Tool name allowlist
disallowedTools	string[]	Tool name denylist
repo	LocalRepoOptions	Clone a GitHub repo and work on a session branch
sandbox	SandboxOptions | boolean	Run in sandbox VM (mutually exclusive with repo)
hooks	GCHooks	Programmatic lifecycle hooks
maxTurns	number	Max agent turns
abortController	AbortController	Cancellation signal
constraints	object	temperature, maxTokens, topP, topK

Message Types

Type	Description	Key Fields
delta	Streaming text/thinking chunk	deltaType, content
assistant	Complete LLM response	content, model, usage, stopReason
tool_use	Tool invocation	toolName, args, toolCallId
tool_result	Tool output	content, isError, toolCallId
system	Lifecycle events	subtype, content, metadata
user	User message (multi-turn)	content

Architecture

my-agent/
├── agent.yaml          # Model, tools, runtime config
├── SOUL.md             # Agent identity & personality
├── RULES.md            # Behavioral rules & constraints
├── DUTIES.md           # Role-specific responsibilities
├── memory/
│   └── MEMORY.md       # Git-committed agent memory
├── tools/
│   └── *.yaml          # Declarative tool definitions
├── skills/
│   └── <name>/
│       ├── SKILL.md    # Skill instructions (YAML frontmatter)
│       └── scripts/    # Skill scripts
├── workflows/
│   └── *.yaml|*.md     # Multi-step workflow definitions
├── agents/
│   └── <name>/         # Sub-agent definitions
├── hooks/
│   └── hooks.yaml      # Lifecycle hook scripts
├── knowledge/
│   └── index.yaml      # Knowledge base entries
├── config/
│   ├── default.yaml    # Default environment config
│   └── <env>.yaml      # Environment overrides
├── examples/
│   └── *.md            # Few-shot examples
└── compliance/
    └── *.yaml          # Compliance & audit config

Agent Manifest (agent.yaml)

spec_version: "0.1.0"
name: my-agent
version: 1.0.0
description: An agent that does things

model:
  preferred: "anthropic:claude-sonnet-4-5-20250929"
  fallback: ["openai:gpt-4o"]
  constraints:
    temperature: 0.7
    max_tokens: 4096

tools: [cli, read, write, memory]

runtime:
  max_turns: 50
  timeout: 120

# Optional
extends: "https://github.com/org/base-agent.git"
skills: [code-review, deploy]
delegation:
  mode: auto
compliance:
  risk_level: medium
  human_in_the_loop: true

Tools

Built-in Tools

Tool	Description
cli	Execute shell commands
read	Read files with pagination
write	Write/create files
memory	Load/save git-committed memory

Declarative Tools

Define tools as YAML in tools/:

# tools/search.yaml
name: search
description: Search the codebase
input_schema:
  properties:
    query:
      type: string
      description: Search query
    path:
      type: string
      description: Directory to search
  required: [query]
implementation:
  script: search.sh
  runtime: sh

The script receives args as JSON on stdin and returns output on stdout.

Hooks

Script-based hooks in hooks/hooks.yaml:

hooks:
  on_session_start:
    - script: validate-env.sh
      description: Check environment is ready
  pre_tool_use:
    - script: audit-tools.sh
      description: Log and gate tool usage
  post_response:
    - script: notify.sh
  on_error:
    - script: alert.sh

Hook scripts receive context as JSON on stdin and return:

{ "action": "allow" }
{ "action": "block", "reason": "Not permitted" }
{ "action": "modify", "args": { "modified": "args" } }

Skills

Skills are composable instruction modules in skills/<name>/:

skills/
  code-review/
    SKILL.md
    scripts/
      lint.sh

---
name: code-review
description: Review code for quality and security
---

# Code Review

When reviewing code:
1. Check for security vulnerabilities
2. Verify error handling
3. Run the lint script for style checks

Invoke via CLI: /skill:code-review Review the auth module

Multi-Model Support

Gitclaw works with any LLM provider supported by pi-ai:

# agent.yaml
model:
  preferred: "anthropic:claude-sonnet-4-5-20250929"
  fallback:
    - "openai:gpt-4o"
    - "google:gemini-2.0-flash"

Supported providers: anthropic, openai, google, xai, groq, mistral, and more.

Inheritance & Composition

Agents can extend base agents:

# agent.yaml
extends: "https://github.com/org/base-agent.git"

# Dependencies
dependencies:
  - name: shared-tools
    source: "https://github.com/org/shared-tools.git"
    version: main
    mount: tools

# Sub-agents
delegation:
  mode: auto

Compliance & Audit

Built-in compliance validation and audit logging:

# agent.yaml
compliance:
  risk_level: high
  human_in_the_loop: true
  data_classification: confidential
  regulatory_frameworks: [SOC2, GDPR]
  recordkeeping:
    audit_logging: true
    retention_days: 90

Audit logs are written to .gitagent/audit.jsonl with full tool invocation traces.

Contributing

Contributions are welcome! Please see CONTRIBUTING.md for guidelines.

License

This project is licensed under the MIT License.