MCP Automations

A production-grade Model Context Protocol server in Python — four LLM-callable tools, two transports, deployed two different ways.

# 30-second proof the server is up:
curl -X POST https://mcp-automations.fly.dev/mcp \
  -H 'Content-Type: application/json' \
  -H 'Accept: application/json, text/event-stream' \
  -d '{"jsonrpc":"2.0","method":"initialize","id":1,
       "params":{"protocolVersion":"2024-11-05",
                 "capabilities":{},
                 "clientInfo":{"name":"curl","version":"1"}}}'

What this is

Most "AI engineer" portfolio projects are applications (a RAG chatbot, an agent that does research). This project is the layer underneath — the typed tools an LLM can call and the transport plumbing that exposes them. MCP is the emerging standard for LLM tool use (~97M monthly SDK downloads as of early 2026); building one — not just consuming one — is the rare skill.

For a deeper look at the design decisions — why two transports, how cost telemetry works, the exception hierarchy, what I'd do differently — see WRITEUP.md.

Related MCP server: production-grade-mcp-agentic-system

Tools

Plus one MCP resource (automations://catalog) and one MCP prompt (daily_brief) — using all three MCP primitives, not just tools.

Every tool returns a typed Pydantic model with per-call token usage and dollar cost attached. Cheap tasks route to Haiku 4.5 ($1/M in, $5/M out), writing-heavy tasks to Sonnet 4.6 ($3/M in, $15/M out).

Connect Claude Desktop to this server

Add one of these to ~/Library/Application Support/Claude/claude_desktop_config.json (Mac) or %APPDATA%/Claude/claude_desktop_config.json (Windows), then restart Claude Desktop.

Option A — local stdio (no network, runs the server as a subprocess):

{
  "mcpServers": {
    "mcp-automations": {
      "command": "python",
      "args": ["-m", "mcp_server.server"],
      "cwd": "/absolute/path/to/mcp-automations",
      "env": {
        "ANTHROPIC_API_KEY": "sk-ant-...",
        "TAVILY_API_KEY": "tvly-..."
      }
    }
  }
}

Option B — remote HTTP (talks to the live Fly server, no local setup):

{
  "mcpServers": {
    "mcp-automations": {
      "url": "https://mcp-automations.fly.dev/mcp",
      "transport": "http"
    }
  }
}

Then ask Claude something like "summarize https://www.paulgraham.com/greatwork.html in 3 bullets" — it'll call summarize_url automatically.

Run locally

pip install -r requirements.txt

# Stdio (for Claude Desktop):
python -m mcp_server.server

# HTTP server (defaults to 0.0.0.0:8765):
MCP_TRANSPORT=http python -m mcp_server.server

# Streamlit playground:
streamlit run playground/app.py

Requires Python 3.13. Needs ANTHROPIC_API_KEY and TAVILY_API_KEY in .env (see .env.example).

Architecture

┌────────────────┐     stdio      ┌──────────────────────┐
│ Claude Desktop ├───────────────►│                      │
└────────────────┘                │                      │
                                  │   mcp_server/        │
┌────────────────┐    HTTP/JSON   │   server.py          │
│ Remote client  ├───────────────►│   (FastMCP)          │
└────────────────┘   (Fly.io)     │                      │
                                  │   4 tools            │
┌────────────────┐  direct call   │   1 resource         │
│ Streamlit UI   ├───────────────►│   1 prompt           │
└────────────────┘                └──────────┬───────────┘
                                             │
                                  ┌──────────┴───────────┐
                                  │ Anthropic + Tavily   │
                                  │ (lazy clients)       │
                                  └──────────────────────┘

The same Python callables back all three entry points. The transport is just a wrapper.

What's in this repo

mcp-automations/
├── mcp_server/
│   ├── server.py        # FastMCP server: 4 tools + 1 resource + 1 prompt
│   ├── models.py        # Pydantic I/O schemas (incl. per-call Cost telemetry)
│   └── exceptions.py    # MCPToolError + UpstreamAPIError / EmptyLLMResponseError / ExtractionError
├── playground/
│   └── app.py           # Streamlit UI with per-session call + spend caps
├── tests/               # offline unit tests (httpx/anthropic/tavily all mocked)
├── Dockerfile           # python:3.13-slim, MCP_TRANSPORT=http for Fly
├── fly.toml             # shared-cpu-1x, 256mb, auto-stop when idle
└── .github/workflows/   # ruff + strict mypy + pytest on every push

Production touches worth noting

• Cost telemetry on every tool response (models.Cost) — token counts and USD attached so a client doesn't have to re-derive it.

• Cost-aware model routing — cheap tasks → Haiku, writing tasks → Sonnet.

• Domain-specific exception hierarchy — UpstreamAPIError, EmptyLLMResponseError, ExtractionError each route differently in logs and the Streamlit UI.

• Two transports, one codebase — MCP_TRANSPORT=stdio|http env switch; HTTP host/port from env so the same image runs on Fly.

• Per-session abuse caps in the playground — 20 calls / $0.50 max per session; backed by a $2/mo hard cap on the Anthropic console.

• Strict mypy + ruff + pytest in CI on every push (.github/workflows/ci.yml).

Why MCP

MCP is transport-agnostic, so one server serves both a local Claude Desktop user (stdio subprocess) and a hosted multi-tenant deployment (HTTPS). It also exposes three primitives that most demos skip:

• Tools — functions the model decides to call (4 of them here)

• Resources — read-only data the client can fetch by URI (automations://catalog returns the tool list as JSON)

• Prompts — server-side templates the user explicitly invokes (daily_brief chains daily_digest + repurpose_content)

Using all three is a signal of reading the spec, not just a quickstart.

Status

✅ Code on GitHub, CI green ✅ Public playground on Streamlit Cloud ✅ Public MCP HTTP server on Fly.io ✅ Cost protection (per-session caps + monthly Anthropic cap) ✅ Real test coverage (23 offline unit tests) ✅ Listed on the Official MCP Registry as io.github.wzltmp/mcp-automations ✅ Long-form writeup of design decisions ✅ Consumed by another agent, not just demoed — langgraph-research-agent's read_node calls this server's summarize_url tool over HTTP (with local fallback if the call fails) 🚧 Demo gif + screenshots (planned) 🚧 n8n self-host via docker-compose (planned)

License

MIT.