claude-code-mcp

It ssounds dumb... but its literally that. Using Claude Code through an MCP server. Its neant for orchestrator agents (like Hermes), giving it fine-grained control over interactive Claude Code sessions running inside tmux.

The core value: a long-running Claude Code REPL stays alive in a tmux pane, and your orchestrator can steer it at any point — send the initial prompt, inject a follow-up message mid-session, interrupt a stuck run, or read back only the final answer (token-efficient by default). Way better than using it through 'claude -p'.

Why an orchestrator needs this

When Hermes (or any multi-agent system) delegates a coding task to Claude Code, it needs more than a one-shot call. It needs to:

Need	This server's tool
Send a prompt, get back only the final answer	session_send
Steer a follow-up into an already-running session	session_send on an existing session
Interrupt a runaway or wrong-direction task	session_interrupt
Check whether Claude is still thinking or idle	session_status / health
Resume a session after a restart without losing history	session_start + --resume
Destroy a session when the task is done	session_destroy

Why not claude -p?

claude -p (print / pipeline mode) exits after one response. There is no channel to inject a follow-up — every turn requires spawning a fresh process and loses accumulated context.

This server keeps the interactive REPL (claude with no flags) alive in a tmux pane so prompts can be injected at any time via tmux key injection, history accumulates inside the running process, and mid-run interrupts (Ctrl-C) work as expected.

Features

• Steer mid-session — session_send injects into an already-running session as naturally as typing; no restart required.
• Send & auto-create — if the session doesn't exist yet it is created transparently; no separate session_start call required.
• Resume across restarts — store a Claude session ID once; all future recreations use claude --resume <id> automatically.
• Working directory persistence — supply working_dir once; recreated sessions start in the same directory.
• Hard interrupt — sends Ctrl-C with no extra text injected.
• Health / state — health reports binary availability and all session states at a glance; session_status drills into one session.
• Token-efficient — default output is the extracted assistant text only; full pane dump is opt-in via raw=True.
• Pure stdlib + mcp SDK — no heavy dependencies.

Requirements

Tool	Version
Python	≥ 3.11
tmux	≥ 3.4
Claude Code	≥ 2.0

Installation

pip install claude-code-mcp

Or with uv:

uv tool install claude-code-mcp

For development:

git clone https://github.com/joschi655/claude-code-mcp
cd claude-code-mcp
pip install -e ".[dev]"

MCP configuration

Claude Desktop / Claude Code (~/.claude/claude_desktop_config.json)

{
  "mcpServers": {
    "claude-code-mcp": {
      "command": "claude-code-mcp"
    }
  }
}

With uvx (no install required)

{
  "mcpServers": {
    "claude-code-mcp": {
      "command": "uvx",
      "args": ["claude-code-mcp"]
    }
  }
}

Hermes / custom MCP client

{
  "mcpServers": {
    "claude-code-mcp": {
      "command": "python",
      "args": ["-m", "claude_code_mcp"]
    }
  }
}

Tools

health()

System health check: returns binary availability and a summary of all known sessions.

{
  "tmux_available": true,
  "claude_available": true,
  "session_count": 1,
  "sessions": [...]
}

---

session_send(name, prompt, timeout?, raw?, claude_session_id?, working_dir?)

Send a prompt to a session — or steer a follow-up into an already-running one — and return Claude's final answer. The session is created automatically if it does not exist.

{ "name": "hermes-main", "prompt": "Now run the tests and fix any failures" }

Returns:

{ "response": "...", "timed_out": false, "raw": false }

• timeout (default 120 s) — give up after this many seconds.
• raw=true — returns the full pane dump instead of the extracted answer (useful for debugging).
• claude_session_id — used when auto-creating: runs claude --resume <id>.
• working_dir — used when auto-creating: starts the session in this directory and stores it in metadata for future recreations.

Steering example — Hermes delegates a task, then decides mid-flight to redirect it:

session_send("hermes-main", "implement feature X")
# … later, while Claude is still idle or about to start a wrong approach …
session_send("hermes-main", "actually focus on the auth module first")

---

session_start(name, claude_session_id?, working_dir?)

Explicitly create a new tmux + Claude Code session, or reconnect to an existing one.

• If a tmux session called name already exists it is returned as-is.
• If metadata for name has a stored claude_session_id, runs claude --resume <id> to restore conversation history.
• working_dir is stored in metadata so future recreations start in the same directory.

{ "name": "hermes-main", "working_dir": "/home/user/myproject" }

Returns SessionInfo with tmux_alive, claude_alive, state, claude_session_id, working_dir.

---

session_interrupt(name)

Send Ctrl-C to abort the current operation. No extra text is sent. Use this when Hermes detects that Claude is going in the wrong direction and wants to redirect without waiting for a full response.

---

session_status(name)

Health check for a single session.

{
  "name": "hermes-main",
  "tmux_alive": true,
  "claude_alive": true,
  "state": "idle",
  "claude_session_id": null,
  "working_dir": "/home/user/myproject"
}

---

session_list()

List all managed sessions (from metadata) plus all live tmux sessions.

---

session_destroy(name)

Kill the tmux session and remove stored metadata.

---

session_set_claude_id(name, claude_session_id)

Associate a Claude session ID with name so future restarts can use --resume. Call this after the session is running and you know the ID (e.g. from /session-id inside Claude Code).

Typical orchestrator workflow

# One-step delegation (session is auto-created)
session_send("hermes-main", "implement feature X and run the tests")
# → creates session if needed, waits for Claude to finish, returns final answer

# Steer a follow-up into the running session
session_send("hermes-main", "the tests are failing — fix the import error first")

# Interrupt if Claude is going the wrong way
session_interrupt("hermes-main")
session_send("hermes-main", "forget the previous approach, use the adapter pattern")

# System overview
health()
# → { tmux_available: true, claude_available: true, session_count: 1, ... }

# Single session check
session_status("hermes-main")
# → { state: "idle", working_dir: "/home/user/myproject", ... }

# Tear down when done
session_destroy("hermes-main")

Session ID and --resume

Claude Code stores conversation history under a session ID. To preserve this across tmux restarts:

1. Run session_send("name", "/session-id") to ask Claude for its ID.
2. Call session_set_claude_id("name", "<id>") to persist it.
3. After a restart, session_start("name") (or session_send) will automatically --resume.

Development

# Unit tests (no tmux/claude required)
pytest tests/test_parser.py tests/test_session_logic.py -v

# Integration tests (requires tmux + claude)
CLAUDE_TMUX_INTEGRATION=1 pytest tests/test_integration.py -v

License

MIT — see LICENSE.