Gitagent

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 • Plugins

Why Gitagent?

Most agent frameworks treat configuration as code scattered across your application. Gitagent 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/gitagent/main/install.sh?$(date +%s)")

This will:

• Install gitagent 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:

# Slim CLI + SDK (recommended in sandboxed/CI environments where supply-chain
# scanners reject larger bundles)
npm install -g @open-gitagent/gitagent

# Add voice mode + web UI (the same web UI install.sh launches at :3333)
npm install -g @open-gitagent/voice

install.sh installs both packages by default. Set GITAGENT_SLIM=1 before the curl-bash to skip voice.

Migrating from 1.x → 2.0

Voice mode lives in @open-gitagent/voice now. The reason: as a single bundle, the package was being blocked by some supply-chain scanners that flagged its 3,800-line dist/voice/ui.html and the unused baileys dependency. Splitting voice out drops the slim-core tarball from ~180 kB to ~85 kB and removes the scanner triggers entirely.

# If you were on v1.x and used voice:
npm install -g @open-gitagent/gitagent@latest @open-gitagent/voice

# If you only use the SDK / non-voice CLI:
npm install -g @open-gitagent/gitagent@latest

The gitagent command and @open-gitagent/gitagent SDK exports are unchanged. gitagent --voice dynamically loads @open-gitagent/voice; without it installed, it prints a one-line install hint and exits cleanly.

Quick Start

Run your first agent in one line:

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

That's it. Gitagent 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:

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

Resume an existing session:

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

Token can come from env instead of --pat:

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

CLI Options

SDK

import { query } from "gitagent";

// 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 Gitagent 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 "gitagent";

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 "gitagent";

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

Message Types

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
├── plugins/
│   └── <name>/         # Local plugins (plugin.yaml + tools/hooks/skills)
├── 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

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

Plugins

Plugins are reusable extensions that can provide tools, hooks, skills, prompts, and memory layers. They follow the same git-native philosophy — a plugin is a directory with a plugin.yaml manifest.

CLI Commands

# Install from git URL
gitagent plugin install https://github.com/org/my-plugin.git

# Install from local path
gitagent plugin install ./path/to/plugin

# Install with options
gitagent plugin install <source> --name custom-name --force --no-enable

# List all discovered plugins
gitagent plugin list

# Enable / disable
gitagent plugin enable my-plugin
gitagent plugin disable my-plugin

# Remove
gitagent plugin remove my-plugin

# Scaffold a new plugin
gitagent plugin init my-plugin

Plugin Manifest (plugin.yaml)

id: my-plugin                    # Required, kebab-case
name: My Plugin
version: 0.1.0
description: What this plugin does
author: Your Name
license: MIT
engine: ">=0.3.0"               # Min gitagent version

provides:
  tools: true                    # Load tools from tools/*.yaml
  skills: true                   # Load skills from skills/
  prompt: prompt.md              # Inject into system prompt
  hooks:
    pre_tool_use:
      - script: hooks/audit.sh
        description: Audit tool calls

config:
  properties:
    api_key:
      type: string
      description: API key
      env: MY_API_KEY            # Env var fallback
    timeout:
      type: number
      default: 30
  required: [api_key]

entry: index.ts                  # Optional programmatic entry point

Plugin Config in agent.yaml

plugins:
  my-plugin:
    enabled: true
    source: https://github.com/org/my-plugin.git  # Auto-install on load
    version: main                                   # Git branch/tag
    config:
      api_key: "${MY_API_KEY}"                      # Supports env interpolation
      timeout: 60

Config resolution priority: agent.yaml config > env var > manifest default.

Discovery Order

Plugins are discovered in this order (first match wins):

1. Local — <agent-dir>/plugins/<name>/
2. Global — ~/.gitagent/plugins/<name>/
3. Installed — <agent-dir>/.gitagent/plugins/<name>/

Programmatic Plugins

Plugins with an entry field in their manifest get a full API:

// index.ts
import type { GitagentPluginApi } from "gitagent";

export async function register(api: GitagentPluginApi) {
  // Register a tool
  api.registerTool({
    name: "search_docs",
    description: "Search documentation",
    inputSchema: {
      properties: { query: { type: "string" } },
      required: ["query"],
    },
    handler: async (args) => {
      const results = await search(args.query);
      return { text: JSON.stringify(results) };
    },
  });

  // Register a lifecycle hook
  api.registerHook("pre_tool_use", async (ctx) => {
    api.logger.info(`Tool called: ${ctx.tool}`);
    return { action: "allow" };
  });

  // Add to system prompt
  api.addPrompt("Always check docs before answering questions.");

  // Register a memory layer
  api.registerMemoryLayer({
    name: "docs-cache",
    path: "memory/docs-cache.md",
    description: "Cached documentation lookups",
  });
}

Available API methods:

Plugin Structure

my-plugin/
├── plugin.yaml          # Manifest (required)
├── tools/               # Declarative tool definitions
│   └── *.yaml
├── hooks/               # Hook scripts
├── skills/              # Skill modules
├── prompt.md            # System prompt addition
└── index.ts             # Programmatic entry point

MCP (Model Context Protocol)

Gitagent is an MCP client: point it at any MCP server and that server's tools are automatically discovered and made available to the agent — no integration code to write. This unlocks the whole ecosystem of ready-made servers (filesystem, GitHub, Postgres, Slack, fetch, …).

Configure servers in agent.yaml

mcp_servers:
  filesystem:                                   # local server over stdio (default)
    command: npx
    args: ["-y", "@modelcontextprotocol/server-filesystem", "/path/to/data"]
    env:
      LOG_LEVEL: "${MCP_LOG_LEVEL}"             # ${VAR} interpolated from the environment
    timeoutMs: 30000                            # connect/list timeout (default 30000)

  analytics:                                    # remote server over Streamable HTTP
    type: http
    url: "https://mcp.example.com/mcp"
    headers:
      Authorization: "Bearer ${ANALYTICS_TOKEN}"

  legacy:                                       # legacy SSE transport (deprecated)
    type: sse
    url: "https://old.example.com/sse"

On startup gitagent connects to each server, lists its tools, and registers them as <server>__<tool> (e.g. filesystem__read_file, analytics__query). Connections are torn down automatically when the session ends.

Use via the SDK

import { query } from "gitagent";

for await (const msg of query({
  prompt: "Summarize last week's signups from the database",
  mcpServers: {
    postgres: {
      command: "npx",
      args: ["-y", "@modelcontextprotocol/server-postgres", process.env.DB_URL!],
    },
  },
})) {
  if (msg.type === "tool_use") console.log(`calling ${msg.toolName}`);
}

SDK mcpServers are merged with any agent.yaml mcp_servers (the SDK value wins on a key collision).

Behavior & guarantees

• Fail-soft: a server that can't start (or times out) is logged and skipped — other servers and built-in tools keep working.
• Namespaced & sanitized: tool names are prefixed with the server name and cleaned to satisfy provider naming rules.
• Pagination: servers that paginate their tool list are fully enumerated.
• Cleanup: stdio servers (child processes) are shut down on every exit path (normal, /quit, Ctrl+C, error).
• Lazy: if no servers are configured, the MCP SDK is never loaded.

Note: v1 supports MCP tools. Resources and prompts are not yet exposed.

Multi-Model Support

Gitagent 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.

Telemetry

Gitagent ships with built-in OpenTelemetry instrumentation. Set OTEL_EXPORTER_OTLP_ENDPOINT and telemetry is on; leave it unset and runtime cost is zero.

Three layers of signals:

1. HTTP-level — @opentelemetry/instrumentation-undici auto-patches fetch/undici, so every LLM provider call (Anthropic, OpenAI, Google, …) gets a client span with URL, status code, and timing.
2. gen_ai.chat spans — emitted on every assistant message_end. Carry gen_ai.system, gen_ai.request.model, gen_ai.usage.input_tokens, gen_ai.usage.output_tokens, gen_ai.response.finish_reasons, and gitagent.cost_usd. Span/metric content never contains the prompt or completion text.
3. gitagent.tool.execute spans — wrap every tool call with tool.name, tool.call_id, tool.status (ok/error), and tool.error_message on failure.

A root gitagent.agent.session span opens at agent construction and closes on every exit path (success, hook-block, SIGINT, error).

CLI usage

Just set the endpoint — no --import flag, no extra install steps:

OTEL_EXPORTER_OTLP_ENDPOINT=http://localhost:4318 gitagent -p "your prompt"

Telemetry is enabled automatically when the endpoint is set and disabled when it is not. To force-disable even when the endpoint is set, pass GITAGENT_OTEL_ENABLED=false.

Environment variables

SDK usage

For programmatic embedders, call initTelemetry explicitly — you control when initialisation happens:

import { initTelemetry, shutdownTelemetry, query } from "gitagent";

await initTelemetry({ serviceName: "my-app" });

for await (const msg of query({ prompt: "hello", model: "anthropic:claude-4-6-sonnet-latest" })) {
  // …
}

await shutdownTelemetry();

OTEL_EXPORTER_OTLP_ENDPOINT and OTEL_EXPORTER_OTLP_HEADERS are read automatically by the OTLP exporter when not supplied programmatically. Pass exporterEndpoint / headers only when you need to override env-based config in code.

Emitted spans

Emitted metrics

Console quickstart (no collector)

Print spans directly to stdout — useful for local debugging:

OTEL_TRACES_EXPORTER=console gitagent -p "test"

Local Jaeger quickstart

docker run --rm -p 16686:16686 -p 4318:4318 jaegertracing/all-in-one:latest

OTEL_EXPORTER_OTLP_ENDPOINT=http://localhost:4318 gitagent -p "test"

# Open http://localhost:16686 → service "gitagent"

Contributing

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

❓ FAQ

General

What is Gitagent? GitAgent (formerly Gitclaw) is a git-native AI agent framework where the agent IS a git repository. Identity, rules, memory, tools, and skills are all version-controlled files, enabling "agents as repos" paradigm.

How does Gitagent differ from other agent frameworks? Unlike frameworks that scatter configuration across application code, Gitagent makes the agent itself a git repo:

• Fork an agent → inherit personality, rules, tools
• Branch → create alternate personality versions
• git log → see agent's memory evolution
• Diff → track rule changes over time

What is the "agents as repos" concept? Your agent lives in a git repository with structured files:

• 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

Installation & Setup

What are the requirements? Node.js 18+ (or 20+ recommended), npm, and git. Install globally with npm install -g @open-gitagent/gitagent (slim CLI + SDK). Add @open-gitagent/voice for voice mode + the web UI.

How do I set up API keys? Run the installer for guided setup:

bash <(curl -fsSL "https://raw.githubusercontent.com/open-gitagent/gitagent/main/install.sh")

Or set manually:

export OPENAI_API_KEY="sk-..."

Which LLM providers are supported?

• OpenAI (GPT-4o, GPT-4o-mini, etc.)
• Anthropic (Claude models via native SDK)
• Any OpenAI-compatible provider

Use --model flag to override: gitagent --model anthropic:claude-sonnet-4-5-20250929

Core Concepts

What is the SDK and how do I use it? The SDK provides programmatic access via query() function that streams agent events:

import { query } from "gitagent";
for await (const msg of query({ prompt: "hello", model: "openai:gpt-4o-mini" })) {
  if (msg.type === "delta") process.stdout.write(msg.content);
}

How do local repo mode sessions work? Clone a GitHub repo, run an agent on it, auto-commit to a session branch:

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

Resume with: gitagent --repo URL --session gitagent/session-xxx "Continue"

What hooks are available? Hooks are lifecycle scripts or programmatic handlers in hooks/ directory. They trigger on agent events like tool execution, session start/end, or memory updates.

Development

How do I create custom tools? Define tools in tools/ directory using declarative YAML format. Each tool specifies name, description, parameters, and execution logic.

How do I add skills? Create skill modules in skills/ directory. Skills are composable and can be imported from installed packages or defined locally.

What telemetry options are available? OpenTelemetry integration for observability:

• Set OTEL_EXPORTER_OTLP_ENDPOINT for auto-enable
• Use OTEL_TRACES_EXPORTER=console for local debugging
• Jaeger quickstart with Docker

Troubleshooting

Why is my agent not responding?

• Check API key is set (OPENAI_API_KEY or equivalent)
• Verify network connectivity to LLM provider
• Use --verbose flag for detailed logs
• Check agent.yaml model configuration

How do I debug agent behavior?

• Use console exporter: OTEL_TRACES_EXPORTER=console gitagent -p "test"
• Check spans in Jaeger: docker run -p 16686:16686 -p 4318:4318 jaegertracing/all-in-one
• Inspect memory/ directory for agent state

Where can I get help?

• GitHub Issues: https://github.com/open-gitagent/gitagent/issues
• Examples: See README SDK section and CLI options
• Contributing: See CONTRIBUTING.md for guidelines

License

This project is licensed under the MIT License.