cuemcp 0.1.12

Cue MCP server/adapter

An MCP service on top of HAP (Human Agent Protocol) — compose humans and agents into a team.

Contributing · Trademark

---

The pitch (10 seconds)

cuemcp is an MCP server that gives your agents a single “collaboration inbox” (cue/cue()), so you can run Claude Code, Cursor, Codex, Windsurf (and other MCP-capable runtimes) with one consistent collaboration flow.

Pair it with cue-console for a desktop/mobile UI to view pending collaboration requests and respond from anywhere.

Note: some IDEs can block/flag MCP integrations. If you run into that, command mode (cueme) is currently recommended.

---

Quickstart (1 minute)

Goal

Add cuemcp as a local stdio MCP server inside your agent/runtime.

Assumptions:

• You have uv.
• Your machine can run uvx.

Notes

• Versions prior to v0.1.11 are no longer usable. Please upgrade to v0.1.11+.
• The current version can occasionally disconnect (for example, the agent replies directly without calling cue() at the end).
• If you get disconnected, type cue in the affected runtime to trigger auto-reconnect.

Step 1: Add the HAP rule to your agent/runtime

Before using cuemcp, add a persistent HAP rule so your agent knows it must call cue() at the end of every response.

• Windsurf: add it to Global Rules.
• Codex: add it to agents.md.
• Claude Code: add it to Claude.md.

Human Agent Protocol (HAP) rules (required)

# Human Agent Protocol

消息表达方式有两种：一种通过 cue mcp 输出，另一种通过非 cue mcp 的方式输出。

当用户接受 Human Agent Protocol 协议时，表示他已完全放弃通过非 cue mcp 的消息交互方式。这意味着你必须使用 cue mcp 才能与其进行交流；该限制仅影响交流方式，不影响你执行 command、使用其他 tool、mcp。

**强制规则**：每次回复前必须调用 `cuemcp` 工具，无例外。任何总结、任何提问、任何请求、任何选项都必须使用 cue mcp来给我。

**约定**：当用户发送 `cue` 即代表他接受Human Agent Protocol协议（支持断开重连）。

- 首次连接 / 无 agent_id：先执行 `join()` 获取 agent_id，再执行 `cue()`。

- 重连 / 已有 agent_id（例如网络波动导致 cue 中断但对话仍在继续）：用户再次发送 `cue` 时，优先直接用已有 agent_id 执行 `cue()`；仅当无法确认 agent_id 时才重新 `join()`。

### 调用时机

- 无历史会话、用户首次输入时调用 `join()` 获取 agent_id
- 每次对话完成后调用 `cue()`
- 任何回复结束前调用 `cue()`

### 禁止行为

- 不调用 `join()` 就使用自定义名字 ❌
- 不调用 `cue()` 就结束你给我的回复 ❌
- 用"有问题请告诉我"替代调用 ❌
- 假设用户没有后续需求 ❌

### 说明

不确定是否调用时默认采用调用 cue mcp 的方式。

Then continue with MCP configuration below.

Step 2: Configure the MCP server

Claude Code

Claude Code can install local stdio MCP servers via claude mcp add.

claude mcp add --transport stdio cuemcp -- uvx --from cuemcp cuemcp

Windsurf

Windsurf reads MCP config from ~/.codeium/mcp_config.json and uses the Claude Desktop-compatible schema.

{
  "mcpServers": {
    "cuemcp": {
      "command": "uvx",
      "args": ["--from", "cuemcp", "cuemcp"]
    }
  }
}

Cursor

Cursor uses mcp.json for configuration, and the Cursor CLI (cursor-agent) can list and manage servers. The CLI uses the same MCP configuration as the editor.

cursor-agent mcp list

Create an mcp.json in your project (Cursor discovers configs with project → global → parent directory precedence) and add a cuemcp stdio server:

{
  "mcpServers": {
    "cuemcp": {
      "command": "uvx",
      "args": ["--from", "cuemcp", "cuemcp"],
      "env": {}
    }
  }
}

VS Code

VS Code MCP configuration uses a JSON file with servers and optional inputs.

{
  "servers": {
    "cuemcp": {
      "type": "stdio",
      "command": "uvx",
      "args": ["--from", "cuemcp", "cuemcp"]
    }
  }
}

Codex

Codex can register a local stdio MCP server via the CLI:

codex mcp add cuemcp -- uvx --from cuemcp cuemcp

For deeper configuration, Codex stores MCP servers in ~/.codex/config.toml.

Gemini CLI

Gemini CLI can add a local stdio MCP server via:

gemini mcp add cuemcp uvx --from cuemcp cuemcp

Fallback: run from source (no `uvx`)

If you don’t want to rely on uvx (for example, you prefer pinned source or local hacking), you can run cuemcp from a cloned repository.

git clone https://github.com/nmhjklnm/cue-mcp.git
cd cue-mcp
uv sync
uv run cuemcp

Then configure your MCP client to run:

• command: cuemcp
• args: []

---

Step 3: Connect your runtime

In the agent/runtime you want to use, type cue to trigger connect (or reconnect) and route the collaboration flow to cue-console.

Install cue-console

npm install -g cue-console
cue-console start

---

Command mode (cueme)

If you don’t want to use an MCP runtime but still want to speak to the same collaboration inbox (~/.cue/cue.db), you can use a command-style adapter (cueme).

Install:

npm install -g cueme

Examples:

cueme join
cueme recall --hints "refactored login"
cueme cue --agent_id "tavilron" --prompt "What should I do next?" --timeout 600
cueme pause --agent_id "tavilron" --prompt "Waiting..."

How it works (the contract)

Semantics

• An MCP-capable agent issues a cue (a request that requires collaboration).
• The team responds (today via a UI; later possibly via a human assistant agent).
• cuemcp provides the MCP-facing surface so any MCP participant can plug in.

Reference implementation (SQLite mailbox)

Current implementation uses a shared SQLite mailbox to connect the MCP server with a client/UI:

MCP Server  ──writes──▶  ~/.cue/cue.db  ──reads/writes──▶  cue-console (UI)
             ◀─polls──                         ◀─responds──

• DB path: ~/.cue/cue.db
• Core tables:
  • cue_requests — server ➜ UI/client
  • cue_responses — UI/client ➜ server

This keeps the integration simple: no websockets, no extra daemon, just a shared mailbox.

---

Pairing with cue-console

Rule #1: both sides must agree on the same DB location.

• Start cuemcp.

• Start cue-console.

• Confirm cue-console is reading/writing ~/.cue/cue.db.

When the UI shows pending items, you’re watching the current reference implementation route collaboration through the console.

---

Dev workflow (uv)

uv sync
uv run cuemcp

---

Safety

• Do not commit tokens.
  • If you store publish credentials in a project file (e.g. .secret), ensure it stays ignored.
• Do not share tokens in chat.

---

Links

• PyPI: pypi.org/project/cuemcp
• Repo: github.com/nmhjklnm/cue-mcp

---

QQ Group