MCP Cross-Model Delegation MCP server

MCP Cross-Model Delegation

English | Français

A security-conscious MCP bridge and research testbed for cross-model LLM delegation.

Status: Experimental / Research Preview (v0.1.0). This is a small, local-first example for studying whether bounded delegation helps an orchestrating model. It is not production-ready and does not claim that delegation improves quality, speed, cost, or token use. The first secondary provider is Gemini; the provider contract permits future implementations without changing the MCP tools.

Why this project exists

An orchestrator sometimes has a bounded task or a long context that a second model could process. The hypothesis is that delegation may reduce orchestrator context load or improve specialized processing, but may also add latency, cost, privacy risk, and error propagation. This repository supplies an inspectable bridge and a measurement plan; it contains no experimental results. See the research roadmap and benchmark schema.

flowchart TD
    A[Orchestrating LLM] --> B[MCP client]
    B --> C[Local MCP delegation server]
    C --> D[Validation and prompt boundary]
    D --> E[DelegationProvider]
    E --> F[Gemini API]
    F --> E --> D --> C --> B
    B --> G[Orchestrator verification and synthesis]

The orchestrator chooses what to send. The MCP server validates and bounds text. GeminiProvider calls Gemini. The orchestrator must verify the returned answer; the bridge does not verify factual accuracy.

MCP tools

Both tools accept text provided by the caller. They do not automatically read Google Drive, email, files, or user applications; execute code; or take actions in those applications. They cannot guarantee correctness. Treat evidence returned by Gemini as a claim to check against the original text.

Example calls with synthetic data:

gemini_delegate_task(task="Calculate 137 × 29 and explain the calculation briefly.")

gemini_extract_findings(
    question="What are the two numbers mentioned and their sum?",
    context="Monday: 12 tickets. Tuesday: 8 tickets."
)

Quick start

Requires Python 3.12+, uv, and a Gemini API key for real calls. Tests need no key.

git clone https://github.com/YOUR_GITHUB_USERNAME/mcp-cross-model-delegation.git
cd mcp-cross-model-delegation
cp .env.example .env
# Edit .env locally and replace GEMINI_API_KEY=replace-me with your own key.
uv sync --extra test
uv run --env-file .env python server.py

The streamable HTTP MCP endpoint is http://127.0.0.1:8000/mcp. Point a local MCP client at it. .env is ignored by Git. Do not paste your key into a command, issue, log, or chat. GEMINI_MODEL defaults to Google's documented stable gemini-3.5-flash-lite; users can choose another supported Gemini model through the environment. Confirm model availability and pricing for your own project.

Local development and Docker

uv sync --extra test
uv run pytest -q
uv run ruff check gateway.py server.py providers tests

docker build -t mcp-cross-model-delegation:0.1.0 .
# Linux host networking example; the MCP listener remains loopback-only.
docker run --rm --network host --env-file .env mcp-cross-model-delegation:0.1.0

The image contains no key. Docker Desktop host networking varies by platform; use a local Python run when host networking is unavailable. The server deliberately refuses MCP_HOST=0.0.0.0. To offer a network endpoint, add suitable client authentication and transport security rather than changing this guard in a shared deployment.

Shortened. The full README is on GitHub.