--- title: "Claude Code MCP server and hooks: connect Claude Code to Cork · Cork" description: "Connect Claude Code to Cork with one machine install, the Claude Code plugin, or the remote MCP server at mcp.corkcli.com: what your agents get, how to check it works, and how to remove it." canonical: "https://corkcli.com/docs/claude-code" last_updated: "2026-10-16" --- # Cork for Claude Code Cork keeps Claude Code agents that work on the same repository, on one machine or across a team's machines, from overwriting each other, and hands each agent what the others already learned. Claude Code gets it through hooks (no extra model calls) and through Cork's MCP server. ## Install for every repo on the machine ``` curl -fsSL https://corkcli.com/install.sh | sh ``` Or, with Node, in any terminal including Windows PowerShell: `npx corkcli`. On Windows the shell script runs in Git Bash, which Claude Code uses too. Either command is safe to run again. What it does: - Downloads Cork, one program, into `~/.cork/bin`, checks it against its SHA-256 before running it, and adds `cork` to your PATH. - Registers the hooks in `~/.claude/settings.json`, the user-level settings, so they apply to every project. Your other settings and hooks are kept. - Adds Cork's MCP tools at user scope as a server named `cork` (the `cork mcp` program, over stdio), in place of any `cork` server already there. - Writes nothing into your repositories: each one is found by its git remote the first time an agent works in it. This is Cork, free and local: no account, and everything stays on the machine. To link the machine to an account run `cork login`; to bring in your team, `cork upgrade`. See [the cork command](https://corkcli.com/docs/cli.md). ## Or install the Claude Code plugin In Claude Code, send these as two separate messages: ``` /plugin marketplace add lithometric/cork /plugin install cork@cork ``` The next session sets Cork up for every repository on the machine. `/cork:status` shows what Cork did on this machine; `/cork:login` links the machine to a Cork account. The plugin is MIT licensed: https://github.com/lithometric/cork ## Or connect the MCP server only ``` claude mcp add --scope user --transport http cork https://mcp.corkcli.com ``` The short form `claude mcp add -t http cork https://mcp.corkcli.com` adds it too. Claude Code signs in through the browser on first use (OAuth 2.1 with PKCE; no API key, headers or client ID). The first Cork call also installs the same machine hooks; if your agent skips that, tell it `set up cork`. Its `setup` tool returns the one install command, which links the machine to your account. Added Cork as a connector on claude.ai? Claude Code loads those connectors too. ## What the agent gets - **A briefing** at session start and with each prompt: the exact signatures, callers and recent changes of the code the prompt is about, so it searches and re-reads less. - **A write gate.** Every write is checked first; one that lands on a teammate's live change is held, and the agent is told who and what, so it can adapt and retry. - **Teammates' changes** on its next step: lines starting with `Δ` in a tool result name what changed since its last step. - **The repo's check** runs detached after a batch of writes, and its verdict arrives with the next tool result. - **Messages** from the other agents in the repo, and `cork message` to send one. - **Landing:** a plain `git push` becomes one step that rebases onto teammates' commits, reruns the tests and pushes. - **MCP tools** such as `get_symbol`, `blast_radius`, `recap` and `remember`. See the [MCP tools reference](https://corkcli.com/docs/mcp-tools.md). ### The hook events Claude Code runs all seven: **SessionStart** indexes the tree and prints the briefing, **UserPromptSubmit** narrows it to what the prompt names, **PreToolUse** is the gate, **PostToolUse** reports writes, **PostToolBatch** runs the repo check, **Stop** reports a usage limit, and **ConfigChange** fires when the settings file changes. A blocked write is refused with an exit code and stderr. Every hook fails open: any error, timeout or missing config allows the operation. ## Check that it works - Start a new Claude Code session: a session loads the hooks when it starts, so one that was already open before you installed runs without them. - A **Cork briefing** arrives in the session's context at start. - Run `cork status` to see whether Cork is running on this machine and what it did in the last day and since install. - On a signed-in workspace, edits, intents and collisions appear on the [dashboard](https://corkcli.com/dashboard) as agents work. ## Remove it ``` cork uninstall ``` This takes Cork's hooks out of `~/.claude/settings.json` and its MCP entry out of Claude Code. Your local data stays in `~/.cork` until you delete it. If you also added Cork as a connector on claude.ai, remove it there (Settings, Connectors). Open sessions keep the hooks they started with until they end, so restart them. ## More - [Codex](https://corkcli.com/docs/codex.md) · [Cursor](https://corkcli.com/docs/cursor.md) · [Gemini CLI](https://corkcli.com/docs/gemini-cli.md) - [The cork command](https://corkcli.com/docs/cli.md) · [MCP tools reference](https://corkcli.com/docs/mcp-tools.md) - [Documentation](https://corkcli.com/docs) · [Support](https://corkcli.com/support)