---
title: "Open-source Claude Code plugin: hooks and skill for multi-agent repos · Cork"
description: "Cork's Claude Code plugin is MIT licensed: hooks that report each read, edit and command, a skill that teaches agents to read teammates' changes, and /cork:login to connect a repo. Install in two messages, no MCP setup."
canonical: "https://corkcli.com/open-source"
last_updated: "2026-11-13"
---

# Cork for Claude Code: the hooks and skill, open source

The plugin that connects Claude Code to Cork is MIT licensed. Install it in two messages and the next session sets Cork up for every repository on this machine: every agent you start is briefed on the code its prompt is about, told when a teammate changes what it depends on, and has its pushes landed in one step.

[github.com/lithometric/cork](https://github.com/lithometric/cork/tree/main/plugin)

## Install

# in Claude Code, one message each

/plugin marketplace add lithometric/cork

/plugin install cork@cork

# optional: link this machine to your Cork account

/cork:login

Nothing is written into your repositories: Cork finds each one from its git remote as you work in it. Free needs no account. `/cork:login` opens a browser to link the machine to your account; the credential stays in `~/.cork/credentials.json` on your machine.

## What is in it

hooks/hooks.json
:   The hooks Claude Code runs: at session start, on each prompt, before each write, after each read, edit and command (failed ones too), and when the session ends.

bin/cork
:   The launcher every hook goes through. It downloads Cork's native hook binary once, pinned by SHA-256, and fails open: if it cannot run, it stays silent and never blocks the agent.

skills/cork
:   A skill that teaches the agent to read what Cork tells it: briefings, lines starting with Δ, blocked writes, landings. And to say in one sentence when a teammate's work changed its plan.

commands/
:   /cork:login connects the repository with a browser sign-in. /cork:status says what it is connected to.

## Hooks, MCP, or both

The hooks deliver everything that saves tokens, with no tool calls: in our team benchmark, agents called Cork's MCP tools less than once each, and every measured saving came through the hooks. The MCP server adds tools an agent can call on purpose (look up any symbol, the blast radius of a change, team memory) and works in any MCP client: Codex, Cursor, claude.ai, ChatGPT.

claude mcp add --transport http cork https://mcp.corkcli.com

## What it sends

- Your code's structure: function and class names, signatures, docstrings, who calls whom, and a hash per line. Parsed on your machine; your source files are not sent.
- What your agents are doing: which files they read and edit, the commands they run (shortened, with secrets and local paths taken out), and when.
- Agents' closing messages and session titles, unless your workspace switches them off, and the messages and notes agents write.
- Your prompts, used once to find the code they are about and then dropped; only a numeric summary of their meaning is kept, for two hours.
- Every request is logged on your machine: `cork audit` shows it. The full list is on [What Cork sends](https://corkcli.com/data).

## What is open, and what is not

The plugin (hooks, launcher, skill, commands) is MIT. The native `cork` program it downloads is built from source in the same repository: its hooks and code parser are MIT, its local engine FSL-1.1-MIT (each version becomes MIT two years after release). Cork's cloud service is closed source. Free covers one person with 10 agent hours a month and any number of repos and agents; a second person is [paid agent hours](https://corkcli.com/pricing), $20 a month per 100, pooled across the workspace.

[View on GitHub](https://github.com/lithometric/cork/tree/main/plugin)[See the benchmark](https://corkcli.com/benchmarks/study-3-teams-easy-to-hard)

This page in HTML: https://corkcli.com/open-source