Anodize Mcp— agent skill

<p align="center"> <picture> <source media="(prefers-color-scheme: dark)" srcset="https://raw.githubusercontent.com/msradam/anodize-mcp/main/assets/logo-dark.svg"> <img alt="AnodizeMCP" src="https://raw.githubusercontent.com/msradam/anodize-mcp/main/assets/logo-light.svg" width="110" height="110"> </picture> </p>

<h1 align="center">AnodizeMCP</h1>

A pure-Python framework for building Model Context Protocol servers, with a FastMCP-compatible API, for platforms where Rust cannot go.

The official MCP SDK and FastMCP depend on pydantic-core, which is compiled Rust with no wheel for many targets. AnodizeMCP implements the same FastMCP-style API using only the standard library plus pure-Python dependencies (its one runtime dependency, uvicorn, is pure Python). The server class is AnodizeMCP, also exported as FastMCP, so a server written here moves to FastMCP by changing one import.

Features

• @mcp.tool / @mcp.resource / @mcp.prompt, with input schemas generated from type hints
• stdio and Streamable HTTP transports (HTTP runs under uvicorn)
• Context for logging, progress, sampling, elicitation, roots, and resource reads
• Bearer and JWT auth, middleware, custom routes, completions, lifespan, pagination
• An in-memory test client, like FastMCP's
• No Rust and no compiled extensions, so it installs where pydantic-core cannot (z/OS, ARMv6, AIX, s390x, ...)
• FastMCP-compatible: 169 of FastMCP's own tests pass against it as a parity gate

Install

pip install anodize-mcp

Requires Python 3.9 or newer. The only runtime dependency is uvicorn (pure Python). Verified on z/OS: both install clean.

Quickstart

from anodize_mcp import AnodizeMCP

mcp = AnodizeMCP("demo", instructions="A small demo server.")

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

@mcp.resource("config://app")
def config() -> str:
    return '{"theme": "dark"}'

@mcp.prompt
def review(code: str) -> str:
    return f"Review this code:\n\n{code}"

if __name__ == "__main__":
    mcp.run()  # stdio transport

Tool input schemas are generated from type hints. Supported types include the primitives, Optional/Union, list/dict/set/tuple, Literal, Enum, dataclasses, TypedDict, and stdlib types (datetime, date, UUID, Decimal). Arguments are validated and coerced at call time. Constraints come from Annotated:

from typing import Annotated
from anodize_mcp import Field

@mcp.tool
def scale(factor: Annotated[float, Field(ge=0, le=10, description="0 to 10")]) -> float:
    return factor * 2

A dataclass return value produces an outputSchema and structuredContent automatically.

Why it exists

The barrier is specific: a Rust-based package with no prebuilt wheel for your platform and no way to build one because there is no Rust toolchain. pydantic-core (under both FastMCP and the official SDK) is the clearest case. AnodizeMCP and its dependencies contain no Rust and no compiled code, so they install where those cannot:

• z/OS (the sharpest case): IBM's Open Enterprise SDK for Python bundles cryptography (3.3.2, pre-Rust) and numpy, but there is no rustc targeting z/OS, so pydantic-core cannot be built or installed. AnodizeMCP and uvicorn install clean.
• Linux on IBM Z (s390x), AIX, Solaris/illumos, the BSDs, Cygwin where prebuilt wheels are often absent (on s390x Linux you can build from source, slowly; AnodizeMCP skips the build).
• Exotic or older CPU architectures: ppc64le, riscv64, ARMv6/v7, mips, sparc.
• WebAssembly (Pyodide, PyScript) and locked-down or air-gapped build environments with no compiler, no network, or a no-Rust policy.

The constraint is specifically Rust, not compiled code: C extensions still build.

Drop-in compatibility with FastMCP

The intended workflow: build your server with AnodizeMCP today on a platform where Rust is unavailable, and if Rust later becomes available, switch to FastMCP by changing one import line.

The class is exported as FastMCP, and the decorator and Context APIs match FastMCP's:

from anodize_mcp import FastMCP, Context        # later: from fastmcp import FastMCP

mcp = FastMCP("demo", instructions="...")

@mcp.tool
async def summarize(text: str, ctx: Context) -> str:
    await ctx.info("summarizing")
    result = await ctx.sample(text, system_prompt="Be concise.")
    return result.text

To stay portable both directions, write FastMCP's async style: async def handlers and await ctx.*. AnodizeMCP's Context methods are awaitable for this reason (they also work without await, as a convenience, but that sync-only form does not port back to FastMCP).

This is checked against the official mcp reference client: the same client driving a FastMCP server and an AnodizeMCP server (identical bodies, only the import differs) sees matching tool descriptions, input schemas (additionalProperties: false, parameter defaults), and structured output (scalar returns wrapped as {"result": value} with an outputSchema, like FastMCP).

What ports unchanged

• FastMCP(name, instructions=..., version=..., lifespan=..., icons=..., website_url=..., on_duplicate=..., mask_error_details=..., auth=...); @mcp.tool, @mcp.resource, @mcp.prompt with name/title/description/annotations/tags; add_tool/add_resource/add_prompt
• @mcp.custom_route(path, methods=...), mcp.add_middleware(...), mcp.list_tools/list_resources/list_prompts/get_tool/get_prompt/call_tool/render_prompt, mcp.disable_tool/enable_tool, mcp.disable(names=..., tags=...)/mcp.enable(...)
• ctx: Context injection; await ctx.debug/info/notice/warning/error(...), ctx.log(message, level=...), report_progress, read_resource, list_resources, list_prompts, get_prompt, get_state/set_state/delete_state, send_notification, sample (result .text), elicit(message, dataclass) (result .action/.data), list_roots; ctx.session_id/client_id/request_id/fastmcp/transport/request_context.lifespan_context/access_token
• Parameter types: primitives, Optional/Union/Literal/Enum, list/dict/set/tuple, datetime/date/UUID/Decimal, dataclasses, TypedDict, `pydantic.BaseModel` (built via the server's own pydantic), and constraints via either AnodizeMCP's Field or `pydantic.Field`/`annotated_types` (Annotated[int, Field(ge=0)] validates)
• Return types: str, numbers, dict, list, dataclasses, pydantic models, bytes, None, content blocks (TextContent, ImageContent, ...), the Image/Audio/File helpers, and ToolResult (including is_error)
• mcp.run(transport="stdio"|"http", host=..., port=...)
• fastmcp.Client in-memory, over stdio, over Streamable HTTP (Client("http://host/mcp")), and via a path to a Python script (Client("server.py"))

What is not implemented as of now (use the alternative)

These are not fundamental limits, just features not built yet.

The other expected difference is the negotiated protocol revision: AnodizeMCP implements 2025-06-18 and negotiates down gracefully if the client offers a newer one.

Conformance against FastMCP's own test suite

As a parity check, FastMCP's tests can be pointed at AnodizeMCP by aliasing fastmcp.FastMCP, fastmcp.Client, and the middleware modules to the AnodizeMCP equivalents before test modules import. The conformance/ directory contains the pytest plugin and instructions.

Against FastMCP 3.4.2, 169 tests pass in full (the CI green gate, covering prompts, resources, tools output schema, and rate-limiting/timing middleware). The broader core suite (tools, resources, prompts, server, middleware, client) reaches 599 of 762 (79%).

The remaining failures fall into two design-boundary categories: tests that assert isinstance(x, mcp.types.*) (AnodizeMCP returns plain dicts to avoid the Rust dependency) and features outside the implemented scope (server mounting, task queues, provider integrations, FunctionTool/Tool.from_function). Both are documented in conformance/README.md.

Protocol coverage

Implements MCP protocol revision 2025-06-18.

Context

A handler receives a Context by declaring a parameter annotated as Context. It is excluded from the input schema and injected at call time.

from anodize_mcp import Context

@mcp.tool
def review(code: str, ctx: Context) -> str:
    ctx.info("starting review")
    result = ctx.sample(f"Review:\n{code}", system_prompt="Be terse.")
    return result.text

Context provides:

• Logging: ctx.debug/info/notice/warning/error(...). Every level is forwarded until the client narrows it with logging/setLevel.
• Progress: ctx.report_progress(progress, total=..., message=...).
• Reading resources: ctx.read_resource(uri).
• Sampling: ctx.sample(messages, system_prompt=..., max_tokens=...) asks the client's LLM. messages is a string, a single message dict, or a list of either.
• Elicitation: ctx.elicit(message, schema) asks the user, where schema is a JSON Schema dict or a dataclass.
• Roots: ctx.list_roots() returns the client's filesystem roots.

sample, elicit, and list_roots are server-to-client requests: the handler blocks until the client responds. They require the client to have declared the matching capability, otherwise they raise an error.

A tool can return a string (text), a dataclass (structured output), or content blocks built directly:

from anodize_mcp import TextContent, ImageContent

@mcp.tool
def render() -> list:
    return [TextContent(text="caption"), ImageContent.from_bytes(png_bytes, "image/png")]

Transports

stdio (default), newline-delimited UTF-8 JSON:

mcp.run()                       # or mcp.run("stdio")
mcp.run("stdio", max_workers=8) # thread pool size for concurrent handlers

Streamable HTTP, a single endpoint (default /mcp):

mcp.run("http", host="127.0.0.1", port=8000)  # serves POST/GET on /mcp

This runs under uvicorn, so production concerns (keep-alive and read timeouts, request size limits, graceful shutdown, signal handling) are handled by uvicorn rather than reimplemented. The uvicorn config matches FastMCP's defaults (timeout_graceful_shutdown=2, lifespan="on") and accepts a passthrough:

mcp.run(
    "http",
    host="0.0.0.0",
    port=8000,
    path="/mcp",
    log_level="info",
    allowed_origins={"localhost", "127.0.0.1"},   # or {"*"} to disable the Origin check
    stateless_http=False,                         # True skips session tracking (alias: stateless)
    uvicorn_config={"timeout_keep_alive": 30},    # any uvicorn.Config setting
)

For your own server (gunicorn, hypercorn, behind a reverse proxy, or mounted in a larger app), get the ASGI app directly. http_app and asgi_app are the same call; http_app matches FastMCP's name:

app = mcp.http_app(path="/mcp", stateless_http=True)
# uvicorn anodize_app:app --host 0.0.0.0 --port 8000

Add ASGI middleware (Starlette Middleware objects, (cls, args, kwargs) tuples, or bare app -> app callables; first item is outermost):

from starlette.middleware import Middleware
from starlette.middleware.cors import CORSMiddleware

app = mcp.http_app(middleware=[Middleware(CORSMiddleware, allow_origins=["*"])])

Mount under an existing Starlette/FastAPI app. The mounted app's lifespan is not driven by the parent automatically, so pass app.lifespan:

from starlette.applications import Starlette
from starlette.routing import Mount

mcp_app = mcp.http_app(path="/")
app = Starlette(routes=[Mount("/mcp", app=mcp_app)], lifespan=mcp_app.lifespan)

event_store / retry_interval (FastMCP's resumable-SSE options) are not implemented.

HTTPS / TLS

Like FastMCP, there is no native TLS layer. Production HTTPS is expected to terminate at the edge (reverse proxy, load balancer, or CDN) with the app speaking plain HTTP internally. For direct in-process TLS, pass uvicorn's own settings through uvicorn_config:

mcp.run("http", port=8443,
        uvicorn_config={"ssl_keyfile": "key.pem", "ssl_certfile": "cert.pem"})

If uvicorn is somehow not installed, run("http", ...) falls back to the standard-library http.server; that fallback does not honor uvicorn_config, so passing one (TLS or otherwise) raises rather than silently serving plaintext. The HTTP transport validates the Origin header (localhost only by default), tracks sessions with Mcp-Session-Id, and serves server-to-client messages (progress, logging, sampling) over a GET SSE stream.

Completions

Register argument completers per prompt or resource template:

@mcp.complete_prompt("review")
def complete(argument: str, value: str) -> list[str]:
    if argument == "language":
        return [x for x in ("python", "rust", "go") if x.startswith(value)]
    return []

A completer may take a third context argument (the already-entered values) and may return a CompletionResult(values=..., total=..., has_more=...) for explicit totals.

Authentication

Authentication applies to the HTTP transport. stdio relies on the operating system process boundary (the server runs under the identity of whoever launched it), so it has no token layer.

The model matches FastMCP: pass a token verifier to the server, the HTTP layer reads Authorization: Bearer <token>, and a handler reads the result with get_access_token() or ctx.access_token. Issuing tokens is left to an external identity provider.

from anodize_mcp import AnodizeMCP, Context, StaticTokenVerifier, get_access_token

mcp = AnodizeMCP(
    "demo",
    auth=StaticTokenVerifier({"dev-token": {"client_id": "cli", "scopes": ["read"]}}),
)

@mcp.tool
def whoami(ctx: Context) -> str:
    token = ctx.access_token            # also: get_access_token()
    return f"{token.client_id} {token.scopes}"

A request with no token gets 401 and a WWW-Authenticate: Bearer header; an invalid token gets 401; a valid token missing a required scope gets 403.

JWTVerifier validates JSON Web Tokens. HS256/384/512 (HMAC) use only the standard library; RS256/384/512 use the cryptography package if it is importable and raise a clear error otherwise.

from anodize_mcp import JWTVerifier

# Symmetric (HS256), standard library only
mcp = AnodizeMCP("demo", auth=JWTVerifier(secret="...", issuer="https://idp", audience="my-api"))

# Asymmetric, keys fetched from the IdP (needs cryptography for RS256)
mcp = AnodizeMCP("demo", auth=JWTVerifier(jwks_uri="https://idp/.well-known/jwks.json",
                                          issuer="https://idp", audience="my-api"))

The verifier is any object with verify_token(token: str) -> AccessToken | None and an optional required_scopes, so a custom verifier (LDAP, RACF, a database lookup) drops in. The OAuth 2.1 authorization-server flow and the hosted-IdP provider wrappers are not implemented as of now; point those at your IdP and verify the tokens here.

Lifespan

Run setup and teardown around the server with lifespan, a context manager whose yielded value is available to every handler. Synchronous and asynchronous context managers both work; synchronous resources are the clean case, an async resource bound to an event loop carries the usual cross-loop caveat since each handler runs on its own loop.

import contextlib

@contextlib.contextmanager
def lifespan(server):
    pool = open_connection_pool()
    try:
        yield {"pool": pool}
    finally:
        pool.close()

mcp = AnodizeMCP("demo", lifespan=lifespan)

@mcp.tool
def query(sql: str, ctx: Context) -> list:
    return ctx.request_context.lifespan_context["pool"].run(sql)

Custom routes

Register handlers at arbitrary HTTP paths for health checks, metrics, or OAuth callbacks. Custom routes bypass the MCP auth and Origin checks (HTTP transport only). A handler returns a Response, a (status, body) tuple, a dict/list (JSON), a str, or bytes.

@mcp.custom_route("/health", methods=["GET"])
def health(request):
    return {"status": "ok"}

The decorator and the handler(request) -> response shape match FastMCP; the request and response objects are AnodizeMCP's own (no Starlette dependency).

Middleware

add_middleware wraps a chain of hooks around request dispatch. Hook names and the (context, call_next) shape match FastMCP. on_message runs for every request; per-operation hooks (on_call_tool, on_read_resource, on_get_prompt, ...) run nested inside for the matching method.

from anodize_mcp import Middleware

class Timing(Middleware):
    async def on_call_tool(self, context, call_next):
        result = await call_next(context)
        return result

mcp.add_middleware(Timing())

Testing with the in-memory client

Client connects to a server with no network in between, the way FastMCP's test client does, so a test exercises the real MCP request path. Pass the server object for an in-process connection, or a command list to launch a subprocess over stdio.

import asyncio
from anodize_mcp import AnodizeMCP, Client

mcp = AnodizeMCP("demo")

@mcp.tool
def add(a: int, b: int) -> int:
    return a + b

async def main():
    async with Client(mcp) as client:                 # in-process
        result = await client.call_tool("add", {"a": 1, "b": 2})
        assert result.data == {"result": 3}

    async with Client(["python", "server.py"]) as client:  # subprocess over stdio
        tools = await client.list_tools()

asyncio.run(main())

The client has list_tools, call_tool, list_resources, read_resource, list_resource_templates, list_prompts, get_prompt, complete, and ping, and follows pagination automatically. Pass sampling_handler, elicitation_handler, and roots to answer the server's ctx.sample/ctx.elicit/ctx.list_roots calls; the client advertises the matching capability and the round-trip runs in process.

Handlers accept either convention: AnodizeMCP's single params dict, or FastMCP's signatures (sampling_handler(messages, params, context), elicitation_handler(message, response_type, params, context), progress_handler(progress, total, message)), chosen by arity. An elicitation handler may return a dict, a dataclass, or a pydantic model (treated as an accept, as FastMCP does), a bare action string ("accept"/"decline"/"cancel"), or a raw {"action": ..., "content": ...} dict.

async with Client(mcp, sampling_handler=lambda params: "the model reply") as client:
    result = await client.call_tool("summarize", {"text": "..."})

Dynamic changes

Registries can change at runtime. Removing an item or calling a notify method broadcasts the corresponding list_changed notification to connected clients:

mcp.remove_tool("old_tool")        # broadcasts notifications/tools/list_changed
mcp.notify_resource_updated(uri)   # to clients subscribed to that uri

Pagination

List endpoints return everything by default, as FastMCP's do. Pass page_size (or FastMCP's list_page_size) to opt in to paging:

mcp = AnodizeMCP("demo", page_size=100)

Clients receive a nextCursor and echo it back. The cursor is opaque. The AnodizeMCP client follows cursors automatically either way.

Development

uv venv && uv pip install -e ".[dev]"
python -m unittest discover -s tests
ruff format . && ruff check . && mypy

The test suite uses only the standard library unittest.

License

MIT. See LICENSE.

Credits

Diode logo by Eucalyp from the Noun Project.