CC-mcp-mimo-bridge

An MCP server that bridges Claude Code and MiMo v2.5 Pro for dual-model collaboration: Claude orchestrates, MiMo implements.

Claude breaks a task into self-contained briefs and delegates them to MiMo through this server. MiMo executes — single-turn, or autonomously across up to 20 agent-loop turns with file/grep/write tools — and returns a structured result. Claude reviews, retries with feedback if needed, and writes the final output to disk.

This is the same pattern Anthropic-internal "subagent" tooling uses, applied to a different model on the other end.

Why this exists

Claude Code is good at decision-making, code review, and long-horizon planning. MiMo v2.5 Pro is cheap and fast at code generation. Pairing them through MCP gives:

• Token-budget offload: Claude no longer spends output tokens on bulk code generation. It writes the brief, MiMo writes the code.
• Two-model review: Claude reviews MiMo's output before any disk write. Bad output is rejected with feedback and retried.
• Auditable: Every agent-loop call dumps a full JSON trace (prompts, tool calls, results, token counts) to .traces/ for inspection.

The BRIEF.md and MIMO_HANDOFF.md files in this repo are the original specs handed to MiMo to implement this very server — kept here as a record of the brief-driven development flow.

Architecture

Claude Code
    │
    │  delegate_to_mimo(brief, …)            single-turn call
    │  agent_delegate_to_mimo(brief, …)      multi-turn agent loop
    │  write_files([{path, content}, …])     batch disk write
    ▼
MCP server (this repo)
    │
    │  Anthropic-compatible API
    ▼
MiMo v2.5 Pro

Inside agent_delegate_to_mimo, MiMo gets its own toolbelt:

| Internal tool | What it does | |---------------|--------------| | read_file | Read a UTF-8 file | | list_dir | List directory entries | | grep | Recursive regex search, capped at 100 matches | | write_file | Write a file (auto-backup on first overwrite per loop) |

The loop runs up to max_turns (default 20). On overflow, MiMo is asked to summarize what it did and what's unfinished. Every loop writes a trace.json to <working_dir>/.traces/<timestamp>_<slug>/ plus per-file backups under backups/.

Repo layout

.
├── BRIEF.md                    # Original spec handed to MiMo (Chinese)
├── MIMO_HANDOFF.md             # Final delivery instructions to MiMo (Chinese)
├── MCP-mimo/                   # Implementation
│   ├── mcp_mimo_server.py      # Server (3 tools)
│   ├── smoke_test.py           # Standalone smoke test
│   ├── requirements.txt
│   └── .gitignore
├── .gitignore
└── README.md                   # this file

Install

git clone https://github.com/<your-account>/CC-mcp-mimo-bridge.git
cd CC-mcp-mimo-bridge/MCP-mimo
pip install -r requirements.txt

Requires Python 3.10+.

Configure

1. Environment variables

| Variable | Required | Default | Description | |--------------------|----------|------------------|----------------------------------------------| | MIMO_BASE_URL | yes | — | MiMo Anthropic-compatible endpoint URL | | MIMO_API_KEY | yes | — | MiMo API key | | MIMO_MODEL | no | mimo-v2.5-pro | Model ID | | MIMO_TIMEOUT_SEC | no | 120 | Request timeout in seconds | | MIMO_MAX_TOKENS | no | 8192 | Max output tokens per request |

2. Register in Claude Code

Add to your ~/.claude.json under mcpServers:

{
  "mcpServers": {
    "mimo-bridge": {
      "command": "python",
      "args": ["/absolute/path/to/CC-mcp-mimo-bridge/MCP-mimo/mcp_mimo_server.py"],
      "env": {
        "MIMO_BASE_URL": "https://your-mimo-endpoint/anthropic",
        "MIMO_API_KEY": "<YOUR_API_KEY>",
        "MIMO_MODEL": "mimo-v2.5-pro",
        "MIMO_TIMEOUT_SEC": "120",
        "MIMO_MAX_TOKENS": "8192"
      }
    }
  }
}

Or via CLI:

claude mcp add mimo-bridge \
  -- python /absolute/path/to/CC-mcp-mimo-bridge/MCP-mimo/mcp_mimo_server.py

(then add the env block manually).

3. Smoke test

cd MCP-mimo
export MIMO_BASE_URL="https://your-mimo-endpoint/anthropic"
export MIMO_API_KEY="your_key_here"
python smoke_test.py

Expected:

[1/5] Basic delegate_to_mimo call ... PASS
[2/5] With context_files ... PASS
[3/5] With retry_context ... PASS
[4/5] SQLite usage_log has entries ... PASS
[5/5] Structured error when config missing ... PASS

All 5 smoke tests passed.

Tool reference

delegate_to_mimo

Single-turn call. Conversation history is accumulated in-process and persists for the lifetime of the MCP server.

| Parameter | Type | Required | Description | |-------------------|--------|----------|--------------------------------------------------------------------------| | brief | string | yes | Self-contained task brief (MiMo can't see Claude's history) | | expected_output | enum | yes | patch / full_file / code_snippet / structured_json | | context_files | array | no | [{path, content?}] — server reads from disk if content is omitted | | constraints | string | no | Hard constraints (e.g. "pure C99, no new deps") | | retry_context | object | no | {previous_output, review_feedback} for retry loop |

Returns:

{
  "status": "ok | failed | timeout",
  "output": "...",
  "format": "code_snippet",
  "tokens_used": {"input": 1234, "output": 567},
  "model": "mimo-v2.5-pro",
  "error": ""
}

agent_delegate_to_mimo

Multi-turn agent loop. MiMo autonomously calls read_file / list_dir / grep / write_file until it produces a final answer or hits max_turns.

| Parameter | Type | Required | Default | Description | |-------------------|---------|----------|---------|----------------------------------------------------------| | brief | string | yes | — | Self-contained task brief | | expected_output | string | yes | — | Output format | | working_dir | string | yes | — | Working directory; trace and backups land here | | context_files | array | no | — | Same shape as above | | constraints | string | no | — | Hard constraints | | max_turns | int | no | 20 | Hard cap on loop iterations | | allow_write | bool | no | true | When false, MiMo gets read-only tools |

Returns:

{
  "status": "ok | max_turns_exceeded | failed",
  "final_output": "...",
  "trace_path": "<working_dir>/.traces/<timestamp>_<slug>/trace.json",
  "files_written": ["..."],
  "files_backed_up": [{"path": "...", "backup": "..."}],
  "turns_used": 7,
  "tokens_used": {"input": 12345, "output": 6789},
  "model": "mimo-v2.5-pro",
  "error": ""
}

write_files

Batch file write helper. Independent of MiMo — useful when Claude already has the content and just wants to commit it through one tool call.

| Parameter | Type | Required | Description | |-----------|-------|----------|------------------------------------------------------| | files | array | yes | [{path, content}, …] — UTF-8, parents auto-created |

Returns: {"written": [...], "failed": [{"path": ..., "error": ...}, ...]}. A single failure does not abort the rest.

Token-usage tracking

Every call (single-turn and agent-loop) is appended to a local SQLite database:

• Linux/macOS: ~/.euterpe/mcp_mimo_usage.db
• Windows: %APPDATA%\Euterpe\mcp_mimo_usage.db

Schema:

usage_log(id, timestamp, brief_hash, input_tokens, output_tokens, model, status)

brief_hash is the first 16 hex chars of sha256(brief) — enough to deduplicate without leaking content.

Troubleshooting

| Symptom | Likely cause | Fix | |-------------------------------------------|-------------------------------|-----------------------------------------------------------| | MIMO_BASE_URL or MIMO_API_KEY not set | Env vars not exported | Set them in ~/.claude.json env block or shell | | auth failed, check MIMO_API_KEY | Invalid or expired API key | Regenerate at the MiMo dashboard | | connection timeout after Ns | Network/endpoint unreachable | Verify MIMO_BASE_URL, raise MIMO_TIMEOUT_SEC | | rate limited, retry later | Too many requests | Wait; retry-after header is surfaced in the error | | upstream error: 5xx | MiMo server issue | Retry; check MiMo status | | prompt too large: ~Nk tokens | Brief + context > ~32k tokens | Trim or split; for big workloads use agent_delegate_to_mimo (it reads files lazily) |

All errors are returned as structured dicts — the server never raises across the MCP boundary.

Design boundaries

What this server does:

• Protocol-layer bridge between Claude Code and MiMo (Anthropic-compatible API).
• Multi-turn agent loop with file/grep/write tools.
• Local token-usage logging.
• Trace dump and auto-backup for every agent-loop call.

What it does not:

• No semantic judgment of MiMo's output (Claude reviews).
• No concurrent dispatch (one task at a time, by design — keeps tracing simple).
• No result caching (every call hits MiMo).
• No prompt-cache optimization.
• No query_usage tool yet (next version).

License

MIT — see LICENSE if present, otherwise treat the MIT terms as the default.