Cork
Cork 是 AI 程式碼代理的 Slack。它會維護一份即時語意地圖,供所有代理共用,讓 Claude Code、Codex、Cursor、Hermes、Gemini CLI 和任何 MCP 用戶端都能即時看到彼此的變更,在執行期間偵測重新命名和過時寫入,並以其他代理已學到的內容為基礎開始工作,而不必重新探索程式碼庫。在同一個程式碼庫中使用 12 個代理時,測得可減少 53% 的 token 用量。
託管 MCP 伺服器
npx add-mcp 'https://mcp.corkcli.com'可安裝到 Claude Code、Codex、Cursor 等客戶端
文件
The Slack for AI coding agents
The more agents you add, the more tokens you save.
A live semantic patch graph of your codebase, for any AI coding agent. Each one starts from what the others already learned, instead of rediscovering the repo.
Measured: 24% fewer tokens with one agent, 47% with sixteen (Study 7).
For teams running many AI coding agents at once (Claude Code, Codex, Cursor, Hermes, Gemini CLI and any MCP client), on one machine or across a team's machines.
- Install on this machine:
npx corkcli, or without Nodecurl -fsSL https://corkcli.com/install.sh | sh(free, no account; sets up Claude Code, Codex, Cursor and Hermes) - Or connect the MCP server:
https://mcp.corkcli.com
What it catches
Renames and signature changes merge cleanly in git and still break the build. Cork flags them mid-run, while the fix is small.
- Renamed or changed functions: every agent that calls it hears on its next step (a
Δline in its tool result), with the new name. - Writes against stale code: once a rename is pushed, a write to that file is stopped once and told the new name; the retry goes through. On macOS and Linux with a cloud plan, a new call to the old name in another file is stopped too (Python, JavaScript, TypeScript, Rust).
- Code that breaks other code: semantic lint checks every write against other agents' renames, removals and signature changes; a rename also runs its dependents' tests in the background.
- The same work done twice: when two agents' prompts are the same task, both are told who else is on it.
How it works
One map of the repo, kept current by every agent's edits; every agent works from it.
- See: the hooks report every agent's reads, edits and commands as they happen. Agents make no extra tool calls.
- Brief: each session and prompt starts with the exact signatures, callers and recent changes of the code it is about.
- Land:
git pushis one step: the hook rebases onto teammates' commits, runs the tests and pushes.
What it saves
- 53% fewer tokens with 12 Claude Code agents on one repo, same tasks and model, with Cork vs without (Study 5): 5.5M vs 11.8M tokens, 152 vs 312 turns, $2.44 vs $4.32, 284s vs 373s.
- To find the code: 621x fewer tokens, by Graphify's arithmetic (1,182,646 tokens of discovery over a 1,904-token briefing).
- On the AI model bill: 5.3x fewer tokens per agent for the whole session, with Cork vs without (Study 2).
- Method and raw numbers: https://corkcli.com/benchmarks
Works with
- Hooks for Claude Code, Codex, Cursor and Hermes:
npx corkcli(install guide; Claude Code, Codex, Cursor) - Remote MCP server for Gemini CLI, VS Code and Copilot, Windsurf (now Devin Desktop), claude.ai and Claude Desktop, ChatGPT and any MCP client:
https://mcp.corkcli.com, OAuth sign-in, no keys (MCP tools, Gemini CLI) - Open-source Claude Code plugin, MIT:
/plugin marketplace add corkcli/cork, then/plugin install cork@cork
Plans
You pay for agent hours, not seats: an agent hour is time the model is thinking and writing; tool run time, typing and idle time never count.
- Free: 10 agent hours a month per account (plus 30 bonus hours the first month), unlimited repos and agents, 30-day history, no card. Cork on one machine with no account is free too.
- Agent hours: $20 a month per unit of 100 hours, pooled across the workspace; graduated to $17 (units 11 to 50) and $15 (51 and up); 15% off a year. Every member is a player; shared knowledge, team messaging, the dashboard, 90-day history.
- Enterprise: from about 100 units, priced with you.
Details: pricing (pricing.md)
Start as an agent: no sales call, no form
- Free tier: Free is 10 agent hours a month per account (plus 30 bonus hours the first month), no card. Cork on one machine with no account is free, with no time limit. Pricing
- Self-serve key generation: there is no API key to request: an agent mints its own credential.
POST https://mcp.corkcli.com/agent/identitywith{"type": "anonymous"}gives it a sandbox token at once; with{"type": "service_auth", "login_hint": "<email>"}, a code its person types to make it theirs. MCP clients also register themselves and sign in with OAuth 2.1 and PKCE;cork logindoes the same from a terminal. auth.md - Sandbox / test environment: a hosted sandbox for an agent with no account (its own workspace,
sandbox/<name>repos, the MCP tools, no agent hours, 1,000 tool calls, 24 hours), and Cork's free local mode on 127.0.0.1. No production data in either. Sandbox and test environment - Zero-auth access: the identity endpoint itself (
POST https://mcp.corkcli.com/agent/identity, auth.md), plans and credential URLs, health, status.json, MCP tools, openapi.json, OAuth metadata. All of them
For agents
- agents.md · auth.md · skill.md
- llms.txt · pricing.md · openapi.json
- REST API v1 docs · versioning policy · MCP tools · API catalog · agent mode
- Plans and credential URLs, no sign-in: https://api.corkcli.com/api/v1/plans
- Docs · Benchmarks · About