lodestar-cli 0.9.1

Agent-native repo orchestration for multi-agent coordination in Git repositories

Lodestar

Agent-native repo orchestration for multi-agent coordination in Git repositories.

Lodestar is a Python CLI tool that enables multiple AI agents (or humans) to work together on tasks in a Git repository without stepping on each other's toes. It provides atomic task claiming with leases, dependency tracking, and inter-agent messaging—all without requiring a central coordinator or human scheduler.

Why Lodestar?

When multiple agents work in the same repository, chaos ensues: duplicate work, merge conflicts, and broken dependencies. Lodestar solves this by providing:

• Task claiming with leases: Agents atomically claim tasks with TTL-based expiration (no daemons required)
• Dependency-aware scheduling: Tasks automatically become available when their dependencies are verified
• Clean Git history: Task definitions live in .lodestar/spec.yaml (committed), while execution state lives in runtime.sqlite (gitignored)
• No central server: Everything runs locally using SQLite for race-free coordination

Key Features

• Multi-Agent Coordination: Multiple agents can work simultaneously without conflicts through atomic task claims
• Lease-Based Claims: Tasks are claimed with time-limited leases that auto-expire, preventing stuck locks
• Two-Plane State Model: Clean separation between task definitions (spec) and execution state (runtime)
• Dependency Tracking: DAG-based task scheduling with automatic readiness detection
• PRD Context Delivery: Tasks carry just-enough product intent from PRDs—frozen excerpts, section refs, and drift detection
• Self-Documenting CLI: Every command supports --json, --schema, and --explain flags for programmatic access
• Progressive Discovery: No-args commands suggest next actions to guide workflows
• Task Messaging: Built-in task thread communication for handoffs and context sharing
• MCP Server Integration: Expose Lodestar as an MCP server for seamless integration with AI assistants like Claude Desktop

Installation

Using uv (recommended)

uv add lodestar-cli

Using pip

pip install lodestar-cli

Using pipx (for global install)

pipx install lodestar-cli

Verify the installation:

lodestar --version

Quick Start

Initialize a repository and complete your first task:

# Initialize Lodestar in your Git repository
lodestar init

# Check repository health
lodestar doctor

# Create a task
lodestar task create \
  --id "TASK-001" \
  --title "Set up project structure" \
  --description "Create initial directory layout" \
  --priority 1 \
  --label feature

# Join as an agent
lodestar agent join
# Output: Registered as agent A1234ABCD

# Find available work
lodestar task next

# Claim a task (prevents other agents from taking it)
lodestar task claim TASK-001 --agent A1234ABCD

# Do your work...
# git add, git commit, etc.

# Mark task as done
lodestar task done TASK-001

# Verify completion (unblocks dependent tasks)
lodestar task verify TASK-001

# Delete a task (soft-delete, preserves in spec.yaml)
lodestar task delete TASK-002

# Check status
lodestar status

For development/testing, use uv run lodestar instead of lodestar to run from source.

MCP Server Integration

Lodestar can be exposed as an MCP (Model Context Protocol) server, allowing AI assistants like Claude Desktop to interact with your Lodestar repository through standardized tools.

Installing MCP Dependencies

The MCP server functionality requires additional dependencies. Install them using:

# Using uv tool (recommended for global installation)
uv tool install 'lodestar-cli[mcp]'

# Or upgrade if already installed
uv tool upgrade 'lodestar-cli[mcp]'

# Using uv in a project
uv add 'lodestar-cli[mcp]'

# Using pip
pip install 'lodestar-cli[mcp]'

# Using pipx (alternative global installation)
pipx install 'lodestar-cli[mcp]'

# For development
uv sync --extra mcp

Starting the MCP Server

The MCP server supports two transport modes: stdio (for Claude Desktop) and HTTP (for multiple agents).

Stdio Transport (Default)

Run the MCP server from within a Lodestar repository:

# Auto-discover repository from current directory
lodestar mcp serve

# Specify repository path explicitly
lodestar mcp serve --repo /path/to/repository

# Enable logging to file
lodestar mcp serve --log-file mcp.log

# Use JSON-formatted logs
lodestar mcp serve --log-file mcp.log --json-logs

The stdio transport is compatible with MCP clients like Claude Desktop.

HTTP Transport (Multi-Agent)

For multiple agents working in parallel coding sessions, use HTTP transport:

# Start HTTP server on default port (8000)
lodestar mcp serve --transport streamable-http

# Specify custom host and port
lodestar mcp serve -t streamable-http --host 127.0.0.1 --port 8080

# With logging
lodestar mcp serve -t streamable-http --log-file mcp-http.log

The HTTP transport enables multiple agents to connect simultaneously, each with their own session.

Configuring with Claude Desktop

Add Lodestar as an MCP server in Claude Desktop's configuration file:

macOS/Linux: ~/Library/Application Support/Claude/claude_desktop_config.json Windows: %APPDATA%\Claude\claude_desktop_config.json

For single agent (stdio):

{
  "mcpServers": {
    "lodestar": {
      "command": "lodestar",
      "args": ["mcp", "serve", "--repo", "/absolute/path/to/your/repository"],
      "env": {}
    }
  }
}

For multiple parallel agents (HTTP):

{
  "mcpServers": {
    "lodestar": {
      "url": "http://127.0.0.1:8000/mcp"
    }
  }
}

If you installed Lodestar with uv in a project, use:

{
  "mcpServers": {
    "lodestar": {
      "command": "uv",
      "args": ["run", "lodestar", "mcp", "serve", "--repo", "/absolute/path/to/your/repository"],
      "env": {}
    }
  }
}

After updating the configuration, restart Claude Desktop. The Lodestar tools will be available for Claude to use when working with your repository.

Available MCP Tools

When connected, Claude can use Lodestar tools to:

• List and query tasks
• Create and update tasks
• Claim and release tasks
• Mark tasks as done or verified
• Send messages between agents
• View dependency graphs
• And more...

All Lodestar CLI functionality is exposed through the MCP interface.

Architecture Overview

Lodestar uses a two-plane state model that separates concerns:

Plane	Purpose	Location	Git Status
Spec Plane	Task definitions, dependencies, acceptance criteria	.lodestar/spec.yaml	Committed
Runtime Plane	Agent state, leases, heartbeats, messages	.lodestar/runtime.sqlite	Gitignored

This separation provides:

• Clean Git history: No noise from lease claims or heartbeats in version control
• Easy resets: Delete runtime.sqlite to start fresh without losing task definitions
• Reproducibility: Same spec produces same task graph on any machine
• Multi-machine support: Each machine has its own runtime state but shares the task definitions

How It Works

┌─────────────────────────────────────────────────────────────┐
│                     Git Repository                          │
│  ┌─────────────────────────────────────────────────────┐    │
│  │ .lodestar/spec.yaml (committed)                     │    │
│  │  - Task definitions                                 │    │
│  │  - Dependencies (DAG)                               │    │
│  │  - Priorities & labels                              │    │
│  └─────────────────────────────────────────────────────┘    │
│                                                              │
│  ┌─────────────────────────────────────────────────────┐    │
│  │ .lodestar/runtime.sqlite (gitignored)               │    │
│  │  - Registered agents                                │    │
│  │  - Active leases (task claims)                      │    │
│  │  - Task status (ready/done/verified)                │    │
│  │  - Inter-agent messages                             │    │
│  └─────────────────────────────────────────────────────┘    │
└─────────────────────────────────────────────────────────────┘

Tasks flow through states: ready → done → verified. A task becomes claimable when it's ready and all its dependencies are verified. Tasks can also be soft-deleted (status set to deleted) to hide them from normal views while preserving history.

PRD Context Delivery

Tasks can carry product intent from a PRD, so executing agents don't need to re-read the full document:

# Get PRD context for a task
lodestar task context TASK-001

# Create a task with PRD context attached
lodestar task create \
    --id "TASK-002" \
    --title "Implement password reset" \
    --prd-source PRD.md \
    --prd-ref "#password-reset" \
    --prd-excerpt "Reset tokens must expire after 15 minutes..."

When claiming a task, context is delivered automatically. If the PRD has changed since task creation, you'll get a drift warning.

Documentation

Full documentation is available in the docs/ directory:

• Getting Started - Installation and first steps
• Concepts - Architecture and core concepts
  • Two-Plane Model
  • Task Lifecycle
  • Lease Mechanics
  • PRD Context Delivery
• CLI Reference - Complete command documentation
• Guides - How-to guides for common workflows
  • Agent Workflow
  • CI Integration

You can also browse the docs locally:

uv run mkdocs serve

Contributing

Contributions are welcome! This project uses Lodestar for its own task management (dogfooding).

Development Setup

# Clone the repository
git clone https://github.com/lodestar-cli/lodestar.git
cd lodestar

# Install dependencies
uv sync

# Run tests
uv run pytest

# Run linting
uv run ruff check src tests
uv run ruff format src tests

# Serve documentation locally
uv run mkdocs serve

Using Lodestar to Contribute

This repository uses Lodestar for task management. To contribute:

1. Run uv run lodestar status to see available work
2. Find tasks with uv run lodestar task next
3. Claim a task before working: uv run lodestar task claim <id> --agent <your-agent-id>
4. Make your changes and commit
5. Mark complete: uv run lodestar task done <id>
6. Verify: uv run lodestar task verify <id>

See CLAUDE.md for detailed development guidelines.

Testing Standards

Before committing, ensure all checks pass:

uv run ruff check src tests          # Linting
uv run ruff format --check src tests # Format check
uv run pytest                        # Tests
uv run mkdocs build                  # Docs build

License

This project is licensed under the MIT License. See the LICENSE file for details.

Links

• PyPI: lodestar-cli
• Repository: github.com/lodestar-cli/lodestar
• Issues: github.com/lodestar-cli/lodestar/issues
• Documentation: docs/