MCP Gateway

Recruiter hook: Tool-selection accuracy 61% → 94%. Same agent, 12× fewer tool tokens.

Production control plane for AI agents using MCP. Registers 15+ downstream servers (filesystem, GitHub, Postgres, fetch, …), semantically filters 200+ tools to top-8 per request, enforces RBAC, and logs every call to Postgres.

Problem

Your company has multiple MCP servers exposing 200+ tools. Dumping every schema into agent context means wrong tool picks half the time and thousands of wasted tokens on definitions never used.

Solution

Agent → FastAPI Gateway → [Semantic Search → RBAC → Rate Limit → Postgres Audit]
                              ↓ top-8 tools
         ┌────────────────────┼────────────────────┐
         ▼                    ▼                    ▼
    filesystem MCP      GitHub MCP           Postgres MCP
    fetch MCP           GitLab (collision)   + extension servers

Features

#	Feature	Module
1	Real MCP servers	skills/mcp_client — filesystem, fetch, GitHub, Postgres via npx
2	Gateway registration	gateway/registry.py — health-check, optional servers, collision namespacing
3	Semantic tool search	skills/llm_adapter — OpenAI text-embedding-3-small, TF-IDF fallback
4	RBAC	config/rbac_policy.yaml — structured deny reasons
5	Audit log	skills/audit_trail — Postgres + arg redaction, JSONL fallback
6	Rate limiting	Token bucket per caller + tool
7	Benchmark	50 queries — naive vs semantic vs role-scoped
8	RBAC benchmark	30 unauthorized calls — 100% blocked

Metrics

• Tool-selection accuracy: 61% → 94% (semantic top-8)
• Tool-definition tokens: 12× reduction
• RBAC: 100% of unauthorized calls blocked

Quick Start

# With Docker (Postgres + gateway + MCP servers)
cp .env.example .env   # add OPENAI_API_KEY, GITHUB_TOKEN (optional)
docker compose up -d

# Local dev
python -m venv .venv && source .venv/bin/activate
pip install -r requirements.txt
python -m gateway.main

Skills (reusable modules)

• skills/mcp_client — connect to downstream MCP servers, health-check
• skills/llm_adapter — OpenAI embeddings for semantic tool search
• skills/audit_trail — Postgres audit store with sensitive arg redaction

Benchmarks

python benchmarks/run_benchmark.py      # 50-query accuracy benchmark
python benchmarks/run_rbac_benchmark.py  # 30 unauthorized calls

Demo GIF

# Terminal 1: start gateway
python -m gateway.main

# Terminal 2: run demo flow (record with ScreenToGif / OBS)
./scripts/demo_flow.ps1

Demo Users

Email	Role	Password
admin@example.com	admin	demo123
engineer@example.com	engineer	demo123
sre@example.com	sre	demo123
readonly@example.com	readonly	demo123

Stack

FastAPI · MCP SDK 2.0 · OpenAI embeddings · Postgres · Docker Compose · OAuth2/JWT

License

MIT