Skip to main content
Glama

cc-tap

MCP License: MIT Python 3.11+

Experimental MCP server that lets Claude Desktop see into and interact with Claude Code Remote Control sessions.

Example: Ask Claude Desktop "what's my Code session working on?" and get a real answer.

cc-tap demo

Why this exists

Claude Desktop and Claude Code are separate worlds. Both build up rich context on your project, but neither can see what the other is doing. A design idea on Desktop that should drive an implementation in Code, or an implementation detail in Code that the Desktop conversation needs; both require you to copy-paste between windows and re-explain context that already exists.

Claude Code's Remote Control lets you drive a session from another device, but that's still you driving the same session. cc-tap is for the other axis: letting a different agent (Claude Desktop, or another Claude Code instance) read and interact with a running CC session through MCP.

Related MCP server: Claude Code MCP Server

Who is this for

  • Developers who use both Claude Desktop and Claude Code

  • Anyone who wants to monitor or interact with CC sessions programmatically

  • Multi-agent workflows where one Claude instance needs to coordinate with another

Quickstart

pip install cc-tap

Requires an active Claude Code login (claude /login) and Remote Control enabled.

Claude Code

claude mcp add cc-tap -- cc_tap

Claude Desktop

Add to claude_desktop_config.json:

{
  "mcpServers": {
    "cc-tap": {
      "command": "cc_tap",
      "args": []
    }
  }
}

Tools

Tool

Description

list_sessions

List all CC sessions (optionally filter by status)

get_session_info

Get details about a specific session

read_session

Read recent conversation from a session

get_session_events

Get raw events, optionally filtered by type

send_message

Send a message to a CC session (fire and forget)

send_and_wait

Send a message and wait for the full response

How it works

cc-tap reads your Claude Code OAuth credentials (from macOS Keychain or ~/.claude/.credentials.json) and talks to the same API that Claude Code's Remote Control web UI uses. No additional authentication needed.

Sessions are accessed via HTTP polling (~1.5s latency). Messages you send appear in the target CC session as if typed by the user.

Limitations

  • Tool approval — the session runtime only picks up approvals via WebSocket (behind Cloudflare bot protection). You can see pending tool requests via send_and_wait, but must approve them in the CC terminal or claude.ai/code web UI.

  • Not real-time — uses HTTP polling, not WebSocket streaming. ~1.5s latency.

  • Undocumented API — uses internal Anthropic endpoints that may change without notice.

  • Local only — reads credentials from the local machine. Can't be deployed as a remote service.

Protocol

See PROTOCOL.md for the reverse-engineered Claude Code Remote session API documentation.

Development

git clone https://github.com/es617/cc-tap.git
cd cc-tap
pip install -e ".[dev,test]"
pre-commit install
pytest

Disclaimer

This project is an experimental research tool for personal and educational use. It interacts with undocumented, internal Anthropic APIs that are not part of any public or supported API surface. These endpoints may change, break, or be removed at any time without notice.

This project is not affiliated with, endorsed by, or supported by Anthropic. Use it at your own risk. The authors assume no responsibility for any consequences of using this tool, including but not limited to account restrictions, data loss, or service disruption.

By using this tool you acknowledge that you are responsible for compliance with Anthropic's Terms of Service and Acceptable Use Policy.

License

MIT

Available Tools

6 tools
get_session_eventsA

Get raw events from a session, optionally filtered by type.

Args: session_id: The session ID. event_types: Comma-separated types to filter (e.g. 'user,assistant'). Empty for all. last_n: Number of recent events to return (default 50).

Inputs: session_id (required), event_types, last_n. Returns: result

get_session_infoA

Get details about a specific Claude Code session.

Args: session_id: The session ID (with or without 'session_' prefix).

Inputs: session_id (required). Returns: result

list_sessionsA

List all Claude Code sessions.

Args: status_filter: Optional filter — 'active' (non-archived), 'running', 'idle', or '' for all.

Inputs: status_filter. Returns: result

read_sessionA

Read recent conversation from a Claude Code session.

Args: session_id: The session ID. last_n: Number of recent events to return (default 20).

Inputs: session_id (required), last_n. Returns: result

send_and_waitA

Send a message to a Claude Code session and wait for the full response.

Polls until the session produces a result event (turn complete), then returns the assistant's response. Use this when you need the other session's answer.

Args: session_id: The session ID to send to. message: The message text to send. timeout: Max seconds to wait for a response (default 120). poll_interval: Seconds between polls (default 1.5).

Inputs: session_id (required), message (required), timeout, poll_interval. Returns: result

send_messageC

Send a message to a Claude Code session.

Args: session_id: The session ID to send to. message: The message text to send.

Inputs: session_id (required), message (required). Returns: result

Tool Schema Changelog

Recent tool additions, removals, and schema changes observed during successful MCP inspections.

Loading tool schema changelog…

    TDQS

    A3.5/5.0

    Scored across 6 tools

    Disambiguation4/5

    Tools are generally distinct: get_session_info, list_sessions, and send_message/send_and_wait have clear purposes. However, get_session_events and read_session both retrieve session content, with descriptions that partially overlap (events vs conversation), which could cause minor confusion.

    Naming Consistency5/5

    All tools use a consistent verb_noun pattern in snake_case, such as get_session_info, list_sessions, send_message. The naming is predictable and follows a clear convention.

    Tool Count4/5

    With 6 tools, the server is well-scoped for managing Claude Code sessions. The count covers essential operations without being excessive, though a tool for archiving or deleting sessions might be expected.

    Completeness3/5

    The tool set covers listing, reading, and sending to sessions, but lacks tools for lifecycle management like creating, archiving, or closing sessions. This leaves some gaps for full session management.

    Maintenance

    ActivityInactive
    ResponsivenessLimited data

    Related MCP Connectors

    Related MCP Servers