🏛️ Production-Grade MCP Server + Agentic System
A reference implementation of an MCP server designed to actually ship
Multi-tenant · Authenticated · Observable · Rate-limited · Cached · Circuit-broken · Governed
接入你的智能体
從發布者安裝。不是 PlugMCP 端點。
遠端 URL 還是本機指令? 兩種安裝方式的差異
This repository is the companion codebase for a long-form blog post that walks through every single component end to end, with every line of code explained in context. Start there if you want to understand the "why" behind the architecture before reading the code.
Most MCP tutorials end with a @tool decorator that returns "hello world". That is fine for a demo. It is not what ships.
This repository is a reference implementation of an MCP server designed to run in production: multi-tenant, authenticated, observable, rate-limited, cached, circuit-broken, and governed. It exposes a company's heterogeneous data layer (Postgres, Elasticsearch, S3, vector DB) to AI agents as a single, secure tool surface, and ships with a four-agent support copilot (Planner → Retriever → Synthesizer → Critic) that uses it end to end.
The codebase is deliberately organised around twelve components that keep showing up on the 3 AM pager when teams skip them. Each one lives in its own module and can be read, replaced, or extended independently.
The complete production-grade system: MCP server dispatch pipeline on the right, four-agent orchestrator on the left, data plane on top, observability on the bottom, identity and governance as crosscutting concerns.
Each diagram below links back to the corresponding section in the blog, where every line of code is walked through in detail.
Postgres + Row-Level Security · Tenant isolation at the DB layer
Dual transport · Stateless session · Middleware chain
OAuth 2.1 · YAML policies · Human-in-the-loop approvals
Three-level hierarchy · Atomic · Composed · Workflow
Circuit breakers · Retry with jitter · ATBA budget allocator
Redis token bucket · Two-tier cache · Stampede lock
OpenTelemetry · Prometheus · Audit logs · One trace ID
Four-agent design · Planner · Retriever · Synthesizer · Critic
End-to-end agent orchestration with one bounded revise loop
git clone https://github.com/FareedKhan-dev/production-grade-mcp-agentic-system.git
cd production-grade-mcp-agentic-system
cp .env.example .env
Edit .env and set at minimum:
ANTHROPIC_API_KEY — for the agent layerATLAS_AUTH_JWKS_URL — your OAuth 2.1 provider's JWKS endpoint (or leave default for dev)docker compose up -d
That brings up the full local environment:
pip install -e .
export ATLAS_MCP_URL=http://localhost:8080
export ATLAS_MCP_TOKEN=dev-token
export ATLAS_TENANT=acme
export ANTHROPIC_API_KEY=sk-ant-...
atlas-copilot "Why was the refund on order o_9002 for CUST-1001 delayed?"
You will see the four agents run end-to-end, the final draft printed with [S1][S2] citations, and a full trace summary including token counts, tool calls, and the run_id that ties back to Jaeger.
Add this to your MCP host config:
{
"mcpServers": {
"production-mcp": {
"type": "http",
"url": "http://localhost:8080/mcp",
"headers": {
"Authorization": "Bearer ${ATLAS_MCP_TOKEN}",
"X-Tenant-Id": "acme"
}
}
}
}
.
├── 📄 README.md
├── 🐳 docker-compose.yml # Full local stack: app + data + observability
├── 🐳 Dockerfile # Two-stage build, non-root runtime
├── 📜 LICENSE
├── 📦 pyproject.toml # Dependencies, dev tools, CLI entry points
├── ⚙️ .env.example # Every setting documented by component
│
├── 🔧 config/ # Runtime configuration (hot-reloadable)
│ ├── http_allowlist.yaml # Per-tenant outbound HTTP allowlist
│ └── policy.yaml # YAML-driven authorization policies
│
├── 🚢 deploy/ # Deployment sidecar configs
│ ├── otel/config.yaml # OpenTelemetry Collector pipeline
│ ├── prometheus/prometheus.yml # Prometheus scrape targets
│ └── sql/init.sql # Schema + RLS policies + seed data
│
├── 📚 docs/ # Deep-dive documentation
│ ├── AGENT_SYSTEM.md # Multi-agent orchestrator internals
│ ├── ARCHITECTURE.md # The 12 components in detail
│ └── DEPLOYMENT.md # K8s, Cloudflare Workers, bare-metal
│
├── 🧠 src/atlas_mcp/ # Main application source
│ ├── config.py # Centralized typed settings
│ ├── server.py # ⚡ Component 1: Transport & dispatch
│ │
│ ├── 🤖 agents/ # Four-agent support copilot
│ │ ├── planner.py # Emits retrieval plan JSON
│ │ ├── retriever.py # Bounded tool-calling loop
│ │ ├── synthesizer.py # Drafts reply with citations
│ │ ├── critic.py # Approves or sends one revise
│ │ ├── orchestrator.py # Wires the four agents together
│ │ ├── mcp_client.py # Thin JSON-RPC MCP client
│ │ ├── memory.py # STM (Redis) + LTM (vector)
│ │ └── cli.py # atlas-copilot CLI entry point
│ │
│ ├── 🔐 auth/ # Components 2 + 3
│ │ ├── oauth.py # JWT + JWKS validation
│ │ ├── middleware.py # Bearer token extraction
│ │ └── policy.py # YAML-driven policy engine
│ │
│ ├── 🛡️ governance/ # Component 12
│ │ ├── tenant.py # Tenant pinning middleware
│ │ └── approval.py # Human-in-the-loop gate
│ │
│ ├── 🔧 tools/ # Components 4 + 6
│ │ ├── registry.py # In-memory tool index + discovery
│ │ ├── base.py # Tool abstract base + metadata
│ │ ├── atomic/ # Level 1: one backend each
│ │ ├── composed/ # Level 2: deterministic chains
│ │ └── workflow/ # Level 3: multi-step procedures
│ │
│ ├── 🔄 reliability/ # Component 7
│ │ ├── circuit_breaker.py # 3-state machine per tool
│ │ ├── retry.py # Exponential backoff + jitter
│ │ └── atba.py # Adaptive Timeout Budget Allocation
│ │
│ ├── 🚦 ratelimit/ # Component 8
│ │ └── limiter.py # Redis token bucket (Lua-atomic)
│ │
│ ├── ⚡ cache/ # Component 9
│ │ └── manager.py # L1 + L2 cache with stampede lock
│ │
│ ├── 🧱 errors/ # Component 10
│ │ └── framework.py # Structured Error Recovery (SERF)
│ │
│ ├── 🔭 observability/ # Component 11
│ │ ├── tracing.py # OpenTelemetry spans
│ │ ├── metrics.py # Prometheus instruments
│ │ └── audit.py # Structured JSONL audit log
│ │
│ └── ✅ validation/ # Component 5
│ └── schemas.py # Tool call envelope
│
└── 🧪 tests/ # Narrow tests, load-bearing properties
├── test_circuit_breaker.py # State machine transitions
├── test_errors.py # SERF wire format + retry semantics
└── test_policy.py # Deny-beats-allow + default-deny
The test suite is deliberately narrow, covering the three load-bearing safety properties:
pip install -e ".[dev]"
pytest -v
test_circuit_breaker.py — state machine transitions, retryable vs deterministic error classificationtest_errors.py — SERF wire format, retry semantics, MCP-level error datatest_policy.py — default-deny, deny-beats-allow, glob matching, PII condition blockingFor running this in an actual production environment (managed Postgres, real OAuth provider, SIEM integration, Kubernetes), see docs/DEPLOYMENT.md.
Key swaps between local dev and production:
docs/ARCHITECTURE.md — The 12 components in depthdocs/AGENT_SYSTEM.md — Multi-agent orchestrator internalsdocs/DEPLOYMENT.md — Production deployment optionsMIT. See LICENSE.
Built with ☕ and a lot of 3 AM debugging
Building a Production-Grade MCP Server Architecture with a Multi-Agent System
工具
尚無工具清單。遠端伺服器會在首次健康檢查後取得工具清單。
常見問題
在本頁複製適用於 Cursor、Claude Desktop、Claude Code 或 VS Code 的 stdio 設定。它使用發布者的安裝方式,而不是 PlugMCP 代理伺服器。
這筆記錄未列出必要的環境變數。
Production Grade MCP Agentic System 透過 stdio 在本機執行。
上次檢查時間:2026-10-27。 Building a Production-Grade MCP Server Architecture with a Multi-Agent System