easy_mcp

Build secure MCP (Model Context Protocol) servers from plain Python functions.

Watch the demo video and browse the server / client code snapshots in docs/.

easy_mcp is FastAPI-for-MCP: declare a function, add a decorator, run a server. Schema generation, validation, authentication, rate limiting, timeouts, structured logging, and sanitized error handling are all built in — and secure by default.

from easy_mcp import MCPServer

server = MCPServer(port=8000)

@server.tool
def add(a: int, b: int) -> int:
    """Add two numbers."""
    return a + b

server.run()

That's a complete, MCP-compliant server. Connect any MCP client to http://127.0.0.1:8000/mcp and the add tool is discoverable and callable — with its JSON schema generated from the type hints and its description taken from the docstring. Prefer a local, launch-on-demand server for a desktop MCP host? Swap the last line for server.run("stdio").

Installation

pip install easy-mcp-kit

The package installs as easy-mcp-kit; the import name is easy_mcp.

Requires Python 3.11+. Only two runtime dependencies: starlette and uvicorn.

Why easy_mcp?

Quickstart tour

Tool registration

# Bare decorator — name, description, and schema are inferred:
@server.tool
def word_count(text: str) -> dict[str, int]:
    """Count words and characters in a text.

    Args:
        text: The text to analyze.
    """
    return {"words": len(text.split()), "characters": len(text)}

# With options:
@server.tool(name="summarize", tags=("stats",), category="math",
             examples=({"arguments": {"values": [1, 2, 3]}},), timeout=5.0)
def summarize_numbers(values: list[float]) -> dict[str, float]:
    """Compute mean/min/max of a list of numbers."""
    ...

# Async tools just work:
@server.tool
async def fetch_status(url: str) -> str:
    """Fetch a status page."""
    ...

# Dynamic registration at runtime:
server.register_tool(my_function, name="late_tool")
server.unregister_tool("late_tool")

Supported parameter types

Anything else is rejected at registration time with a clear error — never at call time. Validation is strict: booleans are not integers, unknown arguments are hard errors, and every violation is reported (not just the first). Following JSON Schema, a number with no fractional part (3.0) is a valid integer; it reaches your function as int, so range(times) never sees a float.

Transports: Streamable HTTP, SSE, or stdio

The same server object serves every transport; nothing else changes.

server.run()          # HTTP on host:port — Streamable HTTP at /mcp, legacy SSE at /sse
server.run("sse")     # legacy HTTP + SSE only
server.run("stdio")   # stdin/stdout — desktop apps, CLI agents, local MCP hosts

Streamable HTTP is the MCP spec's current HTTP transport. One endpoint, /mcp, takes one JSON-RPC message per POST and answers requests in the response body. initialize opens a session whose MCP-Session-Id header the client echoes on later requests; DELETE /mcp ends it, and sessions idle for an hour expire. The same app keeps serving the legacy /sse + /messages endpoints, so older clients connect unchanged. Protocol versions 2024-11-05 through 2025-11-25 are negotiated during initialize.

from easy_mcp import StreamableHTTPTransport

server.run(StreamableHTTPTransport(
    server,
    path="/mcp",                 # the MCP endpoint
    legacy_sse=False,            # stop serving /sse + /messages
    session_idle_timeout=600.0,  # seconds; None keeps sessions until DELETE
))

Browsers get one more check: their Origin header must be allowlisted, which stops DNS-rebinding pages from reaching a server on your machine. Loopback origins are allowed by default; list any web app that should connect:

server = MCPServer(allowed_origins=["https://app.example.com"])  # "*" allows any

Over stdio the MCP host launches your script as a child process and talks JSON-RPC over its pipes. Logs go to stderr, and sys.stdout is redirected to stderr while serving, so a stray print() inside a tool cannot corrupt the protocol stream. Most desktop MCP hosts take a config entry like this:

{
  "mcpServers": {
    "my-tools": {
      "command": "python",
      "args": ["/path/to/server.py"],
      "env": {"EASY_MCP_STDIO_API_KEY": "optional-key-for-protected-tools"}
    }
  }
}

Authentication works the same way as over SSE, except the credential is the EASY_MCP_STDIO_API_KEY environment variable (or StdioTransport(server, api_key=...)) instead of a header. An invalid key fails at startup rather than silently downgrading to anonymous access.

Authentication and per-tool permissions

from easy_mcp import APIKeyAuth, MCPServer

auth = APIKeyAuth({
    "long-random-admin-key...": "*",          # all scopes
    "long-random-viewer-key..": ["reports"],  # specific scopes
})
# Or keep keys out of code entirely:
# auth = APIKeyAuth.from_env()   # reads EASY_MCP_API_KEYS="key1:*;key2:reports|stats"

server = MCPServer(port=8000, auth=auth)

@server.tool
def public_tool() -> str:
    """Anyone can call this."""

@server.tool(requires_auth=True)
def protected_tool() -> str:
    """Any authenticated client can call this."""

@server.tool(scopes=("admin",))
def admin_tool() -> str:
    """Only keys holding the 'admin' scope can call this."""

Clients authenticate with Authorization: Bearer <key> or X-API-Key. Protected tools are invisible to clients that cannot call them — they are omitted from tools/list and reported as unknown on tools/call, so unauthorized clients cannot even enumerate them.

Rate limiting, payload caps, timeouts, session limits

server = MCPServer(
    port=8000,
    rate_limit_per_minute=120,     # per client; None disables
    max_request_bytes=1_048_576,   # enforced while reading the body
    default_timeout=30.0,          # per tool call; override per tool
)

@server.tool(timeout=2.0, max_calls_per_session=5)
async def expensive(query: str) -> str:
    """A tool with its own timeout and a per-session usage cap."""

Clients can also cancel long-running calls with the standard MCP notifications/cancelled message.

Error handling

In debug=True mode (development only) clients receive full tracebacks. The error_id in production responses matches the server-side log entry that contains the real traceback, so you can correlate without leaking internals.

Structured logging and audit trail

All logs are single-line JSON on stderr. Every tool call is audited with the tool name, client id, duration, and outcome — never with API keys (only SHA-256 fingerprints ever appear):

{"timestamp": "2026-07-20T12:00:00.000Z", "level": "INFO", "logger": "easy_mcp.audit",
 "message": "tool_call", "event": {"type": "tool_call", "tool": "add",
 "client_id": "3f9c2a71b04d", "duration_ms": 0.42, "status": "ok"}}

Connecting a client

# MCP Inspector (interactive UI):
npx @modelcontextprotocol/inspector      # Streamable HTTP, http://127.0.0.1:8000/mcp

From code, any MCP client SDK works. With the official Python SDK (v2):

import asyncio

from mcp import Client


async def main() -> None:
    async with Client("http://127.0.0.1:8000/mcp") as client:
        result = await client.call_tool("add", {"a": 2, "b": 3})
        print(result.content[0].text)  # 5


asyncio.run(main())

Stdio servers are started by the MCP host itself, from a config entry like the one in Transports.

Ready-made connectors

Two servers ship with the package. They are built on the same @server.tool decorator you use, so everything above (validation, scopes, rate limits, timeouts, sanitized errors, audit log) applies to them unchanged.

pip install "easy-mcp-kit[postgres]"   # the GitHub connector needs no extra

GITHUB_TOKEN=github_pat_... easy-mcp-github --transport stdio
DATABASE_URL=postgresql://user:pass@host/db easy-mcp-postgres --port 8011

Both take --transport {http,sse,stdio}, --host, --port, --rate-limit and --debug, and load API keys from EASY_MCP_API_KEYS when it is set. python -m easy_mcp.connectors.github and ... .postgres work as well, and each module's build_server(...) returns a normal MCPServer for embedding.

GitHub is read-only by default. --allow-write registers create_issue and comment_on_issue, which are gated by the github:write scope: only a client presenting a key that holds it can see or call them, and starting with --allow-write but no keys is refused. The token is sent only to GITHUB_API_URL (default https://api.github.com) and never appears in logs or errors. Use a fine-grained token scoped to the repositories you need.

export EASY_MCP_API_KEYS="a-long-random-key:github:write"   # key : scope
easy-mcp-github --allow-write --transport stdio             # stdio: EASY_MCP_STDIO_API_KEY=a-long-random-key

Postgres runs every statement in a READ ONLY transaction (default_transaction_read_only=on is set for the session, so a query cannot turn it off) with a statement timeout (--statement-timeout, default 10 s) and a row cap (--max-rows, default 500; query returns truncated: true when it hit the cap). Writes are rejected by the database itself, not by parsing SQL. Still connect with a dedicated role holding only SELECT grants: a read-only transaction does not stop side-effecting functions that role is allowed to call.

Architecture

easy_mcp/
├── server.py        MCPServer: registration, dispatch, execution, lifecycle
├── decorators.py    @tool machinery, ToolDefinition, thread-safe registry
├── schema.py        type hints → JSON Schema; docstring parsing; validation
├── security/
│   ├── auth.py      APIKeyAuth (constant-time), scopes, visibility rules
│   └── ratelimit.py sliding-window per-client rate limiter
├── transport/
│   ├── base.py      Transport ABC + ClientContext
│   ├── _http.py     shared HTTP plumbing: Origin allowlist, credentials, uvicorn
│   ├── streamable_http.py  Streamable HTTP transport (/mcp, sessions)
│   ├── sse.py       legacy HTTP + SSE transport (Starlette/uvicorn)
│   └── stdio.py     stdin/stdout transport (desktop MCP hosts, local agents)
├── connectors/
│   ├── _cli.py      shared --transport/--host/--port options
│   ├── github.py    GitHub connector (stdlib HTTP; read-only unless --allow-write)
│   └── postgres.py  Postgres connector (psycopg; READ ONLY, timeout, row cap)
├── protocol.py      supported MCP protocol versions + negotiation
├── exceptions.py    error hierarchy + stable JSON-RPC error codes
└── logging.py       JSON logs + audit trail

The dispatcher (MCPServer.dispatch) is transport-independent: it takes one decoded JSON-RPC message plus a ClientContext and returns the response. Transports only resolve credentials, cap payload sizes, and move bytes — so Streamable HTTP, SSE, and stdio all get the same checks, and a future WebSocket transport cannot silently bypass one.

Determinism: tool listings are sorted, JSON output uses sorted keys, and identical inputs produce byte-identical responses — useful for reproducible agent runs and caching.

Performance notes: sync tools run in a worker thread pool so they never block the event loop; async tools run natively. Schema validation is a small hand-written walker (no dependency, ~microseconds for typical payloads). The per-message overhead is dominated by JSON encode/decode; for large results prefer returning compact structures over huge strings.

Production deployment

• Run behind TLS (reverse proxy such as Caddy/nginx) — API keys travel in headers.
• Load keys from the environment (APIKeyAuth.from_env()), never hardcode them.
• Keep debug=False; it is the only thing standing between clients and tracebacks.
• Browser-based clients on other origins must be listed in allowed_origins.
• For multiple workers: uvicorn "myapp:server.build_app" --factory won't share sessions across processes — run one process, or route each MCP-Session-Id to the same worker.
• Read SECURITY.md before exposing a server beyond localhost.

Development

pip install -e .[dev]
pytest            # 100+ tests: schema, dispatch, security, Streamable HTTP, SSE, stdio
ruff check .
mypy easy_mcp

CI runs the same three commands on Python 3.11 through 3.14. Releases are listed in CHANGELOG.md.

License

MIT — see LICENSE.