AgentShare

Share skills and context across AI coding agents (Claude Code, Cursor, Windsurf).

AgentShare gives your AI agents shared memory — when one agent finishes work, the next one picks up where it left off. It also provides a skills registry so you can write reusable instruction snippets once and scaffold them into any project for any platform.

Quick Start

pip install agentshare

# Register the MCP server + inject agent rules into all detected platforms
agentshare mcp init --global

# Restart your AI agents to pick up the changes

That's it. Your agents will now automatically:

• Ask if you want them to fetch prior context for the project
• Use MCP to fetch prior context when you agree
• Save summaries of their work for future agents (write_session)

agentshare mcp init --global also installs an agentshare-cli skill into each detected platform's global skill directory plus ~/.agents/skills. This skill teaches agents how to install and use the AgentShare CLI.

AgentShare also nudges agents to check recent MCP sessions first and only read files when the context is insufficient.

How It Works

AgentShare has two core features:

1. Cross-Agent Context Sharing (MCP Server)

An MCP server exposes four tools to your agents:

Sessions are stored in a local SQLite database (~/.agentshare/context.db) with FTS5 full-text search.

2. Skills Registry

Skills are reusable Markdown instruction files (with YAML frontmatter) that you manage globally and scaffold into projects per-platform.

# Create a skill
agentshare skills create code-review --description "Code review checklist" --category workflows

# Edit it
# ~/.agentshare/skills/workflows/code-review/SKILL.md

# Scaffold into a project for all platforms
agentshare init skills --path ./my-project --all-platforms

Supported Platforms

Platforms are auto-detected based on the presence of their config directories.

CLI Reference

agentshare --version              Show version

agentshare mcp init --global      Register MCP server + inject agent rules + install CLI skill globally
agentshare mcp init               Write .mcp.json to current project (local install)
agentshare mcp serve              Start MCP server (used internally by platforms)
agentshare mcp remove             Remove MCP config + rules + CLI skill from all platforms

agentshare skills list            List all registered skills
agentshare skills add <path>      Import a skill directory
agentshare skills remove <name>   Remove a skill
agentshare skills create <name>   Create a new skill  [-d description] [-c category]

agentshare init skills            Scaffold skills into a project
                                  [--path] [--platform] [--all-platforms] [--category]

Development

git clone https://github.com/devashar13/agentshare.git
cd agentshare
uv venv && source .venv/bin/activate
uv pip install ".[dev]"

# Run tests
uv run pytest -v

Note: After making code changes, re-run uv pip install . to pick them up.

Requires Python 3.11+.

Architecture

~/.agentshare/
  skills/              # Global skills registry
    <category>/<name>/SKILL.md
  context.db           # SQLite + FTS5 session store

src/agentshare/
  cli.py               # Typer CLI app
  config.py            # Paths, platform detection
  context/
    models.py          # Session model (Pydantic)
    store.py           # SQLite CRUD + full-text search
  mcp/
    server.py          # FastMCP server (4 tools)
    installer.py       # Platform config + rules injection
  skills/
    registry.py        # Skill CRUD
    scaffold.py        # Copy skills into project dirs

License

MIT