feat: implement mcp frontend

This commit is contained in:
hh
2026-05-20 09:59:47 +02:00
parent 99a30f256d
commit 7970d4be9b
7 changed files with 902 additions and 24 deletions
+27 -9
View File
@@ -16,8 +16,11 @@ from fastmcp import Client
from fastmcp.client.transports import StdioTransport, StreamableHttpTransport
from fastmcp.server import create_proxy
from beaver_gateway.mcp.lenient import LenientStdioTransport
if TYPE_CHECKING:
from fastmcp import FastMCP
from fastmcp.client.transports.base import ClientTransport
from beaver_gateway.mcp.types import HttpMcp, StdioMcp
@@ -27,22 +30,37 @@ def build_stdio_proxy(spec: StdioMcp) -> FastMCP:
``spec.command`` is a non-empty tuple; the first element is the
executable and the rest are CLI args. ``StdioTransport`` keeps the
subprocess alive across calls.
subprocess alive across calls. When ``spec.lenient`` is set we
swap in :class:`LenientStdioTransport`, which silently drops
non-JSON-RPC stdout lines instead of forwarding them as exceptions
(lets us ingest MCPs that print ``Processing...`` chatter on stdout
without leaking warnings to downstream UIs).
"""
if not spec.command:
msg = f"stdio MCP {spec.name!r} has empty command"
raise ValueError(msg)
command, *args = spec.command
transport = StdioTransport(
command=command,
args=list(args),
env=spec.env,
cwd=str(spec.cwd) if spec.cwd is not None else None,
)
cwd = str(spec.cwd) if spec.cwd is not None else None
transport: ClientTransport
if spec.lenient:
transport = LenientStdioTransport(
command=command, args=list(args), env=spec.env, cwd=cwd
)
else:
transport = StdioTransport(
command=command, args=list(args), env=spec.env, cwd=cwd
)
return create_proxy(Client(transport, name=spec.name))
def build_http_proxy(spec: HttpMcp) -> FastMCP:
"""Wrap a remote streamable-HTTP MCP into a mountable ``FastMCPProxy``."""
transport = StreamableHttpTransport(url=spec.url, auth=spec.auth)
"""Wrap a remote streamable-HTTP MCP into a mountable ``FastMCPProxy``.
``spec.headers`` are folded into the upstream client; pass per-call
auth (e.g. ``X-Api-Key``) here rather than as ``auth`` when the
upstream rejects Bearer headers.
"""
transport = StreamableHttpTransport(
url=spec.url, auth=spec.auth, headers=spec.headers
)
return create_proxy(Client(transport, name=spec.name))
+60 -9
View File
@@ -1,4 +1,4 @@
"""Internal MCP aggregator — one ASGI app, N FastMCP namespaces.
"""Internal MCP aggregator — one ASGI app, N FastMCP namespaces + ``all``.
Each ``McpServer`` declared in the user's config becomes its own
``FastMCP`` instance (regular for ``python_tool``, ``FastMCPProxy`` for
@@ -9,9 +9,18 @@ namespace via loopback as a distinct MCP server URL — preserving
per-domain framing while costing only one process worth of RAM
(PRD §6).
Phase 3 adds ``/mcp/all/``: a single FastMCP whose tools are the union
of every namespace's tools, prefixed by FastMCP's ``namespace_<tool>``
convention (e.g. ``time_current_time``). It's the escape-hatch for
clients that can only configure one MCP server — discouraged for tool-
heavy setups (PRD §6 cites the ~95%→~71% tool-selection drop on flat
namespaces) but real and reachable.
The aggregator returns both the app and a ``{name: url}`` map; Phase
2.2's ``ClaudeCodeBackendAdapter`` plugs the map directly into
``BackendOptions.mcp_servers``.
``BackendOptions.mcp_servers``. ``/mcp/all/`` is NOT included in that
map — claude-code-agents always get per-domain framing; only the
external MCP frontend (Phase 3.1) reverse-proxies the flat endpoint.
"""
from __future__ import annotations
@@ -19,6 +28,7 @@ from __future__ import annotations
from contextlib import AsyncExitStack, asynccontextmanager
from typing import TYPE_CHECKING
from fastmcp import FastMCP
from starlette.applications import Starlette
from starlette.routing import Mount
@@ -29,30 +39,48 @@ from beaver_gateway.mcp.wrap import build_python_tool_server
if TYPE_CHECKING:
from collections.abc import AsyncIterator, Iterable
from fastmcp import FastMCP
from beaver_gateway.mcp.types import McpServerT
ALL_NAMESPACE = "all"
"""URL segment for the flat-namespace aggregator (``/mcp/all/``)."""
def build_internal_app(
mcps: Iterable[McpServerT], *, host: str, port: int
) -> tuple[Starlette, dict[str, str]]:
"""Build the aggregator ``Starlette`` app and the ``{name: url}`` map.
"""Build the aggregator ``Starlette`` app and the per-namespace URL map.
``host``/``port`` only flavour the URL strings handed back — actually
listening on them is the caller's job (``cli.main`` runs a uvicorn
server in a TaskGroup). We accept the address here so callers don't
have to format the URLs themselves and risk drifting from the
``/mcp/<name>`` convention.
Returns a map of ``{namespace: url}`` for the per-domain endpoints
only (claude-code's MCP routing expects per-domain framing). The
``/mcp/all/`` bundle endpoint exists on the same app but is
intentionally omitted from the URL map — it's only meaningful to
external clients via the MCP frontend, not to claude-code.
"""
servers: dict[str, FastMCP] = {spec.name: _build_server(spec) for spec in mcps}
child_apps = [s.http_app(transport="http", path="/") for s in servers.values()]
child_apps = {
name: s.http_app(transport="http", path="/")
for name, s in servers.items()
}
routes = [
Mount(f"/mcp/{name}", app=app)
for name, app in zip(servers, child_apps, strict=True)
Mount(f"/mcp/{name}", app=app) for name, app in child_apps.items()
]
# /mcp/all — flat-namespace bundle. Skip when there's nothing to
# bundle so we don't pay for an empty session manager lifecycle.
all_app = None
if servers:
all_server = _build_all_server(servers)
all_app = all_server.http_app(transport="http", path="/")
routes.append(Mount(f"/mcp/{ALL_NAMESPACE}", app=all_app))
@asynccontextmanager
async def lifespan(_parent: Starlette) -> AsyncIterator[None]:
# Each FastMCP http_app stores its session manager init in its
@@ -61,8 +89,12 @@ def build_internal_app(
# children come up together and unwind in reverse order on
# shutdown.
async with AsyncExitStack() as stack:
for child in child_apps:
for child in child_apps.values():
await stack.enter_async_context(child.router.lifespan_context(child))
if all_app is not None:
await stack.enter_async_context(
all_app.router.lifespan_context(all_app)
)
yield
app = Starlette(routes=routes, lifespan=lifespan)
@@ -85,3 +117,22 @@ def _build_server(spec: McpServerT) -> FastMCP:
# type-narrowing honest if a new variant lands without updates here.
msg = f"unsupported McpServer variant: {type(spec).__name__}"
raise TypeError(msg)
def _build_all_server(children: dict[str, FastMCP]) -> FastMCP:
"""Compose a single FastMCP whose tools union every namespace's tools.
FastMCP's ``mount(namespace=...)`` namespaces every tool as
``<namespace>_<original_name>``, so the flat endpoint becomes
``time_current_time``, ``calendar_event_create``, etc.
Mounted children are accessed in-memory via ``FastMCPProvider`` —
requests don't bounce through the child's own ``http_app``, so this
aggregator has its own independent streamable-HTTP session manager
and lifespan, and the per-namespace ``/mcp/<name>/`` mounts keep
working unchanged.
"""
parent = FastMCP(name="beaver-gateway-all")
for name, child in children.items():
parent.mount(child, namespace=name)
return parent
+234
View File
@@ -0,0 +1,234 @@
"""Tolerant transports for upstream MCPs that don't strictly speak JSON-RPC.
Some real-world MCP servers print non-JSON chatter to stdout before / between
their actual JSON-RPC frames (``Processing...``, banners, dependency-load
messages, etc.). The reference ``mcp.client.stdio.stdio_client`` parses every
stdout line as JSON-RPC and ships any parse failure downstream as an
exception, which the MCP ``ClientSession`` then logs as a warning that bleeds
into client UIs (Cursor, Cline) when they connect through us.
``LenientStdioTransport`` re-implements the stdio-client wiring with one
behavioural change: lines that don't parse as JSON-RPC are *silently
dropped* (one ``DEBUG`` log entry, no exception forwarded). Downstream
consumers see only valid messages. We keep the rest of the contract identical
to the reference client, including the spec-mandated graceful shutdown
sequence (close stdin → wait → SIGTERM → SIGKILL).
The transport plugs into ``fastmcp.server.create_proxy`` just like
``StdioTransport`` does, so the rest of the aggregator doesn't need to
know which flavour it got.
"""
from __future__ import annotations
import contextlib
import logging
import sys
from typing import TYPE_CHECKING
import anyio
import anyio.lowlevel
from anyio.streams.text import TextReceiveStream
from fastmcp.client.transports.base import ClientTransport
from mcp import ClientSession, types
from mcp.client.stdio import (
PROCESS_TERMINATION_TIMEOUT,
StdioServerParameters,
_create_platform_compatible_process,
_get_executable_command,
_terminate_process_tree,
)
from mcp.shared.message import SessionMessage
if TYPE_CHECKING:
from collections.abc import AsyncIterator
from typing import TextIO, Unpack
from anyio.streams.memory import MemoryObjectReceiveStream, MemoryObjectSendStream
from fastmcp.client.transports.base import SessionKwargs
_log = logging.getLogger("beaver_gateway.mcp.lenient")
class LenientStdioTransport(ClientTransport):
"""Stdio transport that tolerates non-JSON-RPC stdout noise.
Behaves like :class:`fastmcp.client.transports.StdioTransport` from the
consumer's perspective: one ``ClientSession`` per ``connect_session``
block, subprocess scoped to that block. We don't replicate the upstream
``keep_alive`` flag because the only caller (``create_proxy``) opens
the session lazily on first request and keeps it open for the lifetime
of the proxy.
"""
def __init__(
self,
command: str,
args: list[str],
env: dict[str, str] | None = None,
cwd: str | None = None,
log_file: TextIO | None = None,
) -> None:
self.command = command
self.args = args
self.env = env
self.cwd = cwd
# TextIO only — pre-open Path callers themselves. The upstream
# ``StdioTransport`` opens Path for you, but we keep this thin so
# the type contract stays narrow and easy to validate.
self.log_file = log_file
@contextlib.asynccontextmanager
async def connect_session(
self, **session_kwargs: Unpack[SessionKwargs]
) -> AsyncIterator[ClientSession]:
errlog: TextIO = self.log_file if self.log_file is not None else sys.stderr
async with _lenient_stdio_client(
StdioServerParameters(
command=self.command,
args=self.args,
env=self.env,
cwd=self.cwd,
),
errlog=errlog,
) as (read_stream, write_stream), ClientSession(
read_stream, write_stream, **session_kwargs
) as session:
yield session
def __repr__(self) -> str:
return (
f"<LenientStdioTransport(command={self.command!r}, args={self.args!r})>"
)
@contextlib.asynccontextmanager
async def _lenient_stdio_client( # noqa: PLR0915 — mirrors mcp.client.stdio.stdio_client
server: StdioServerParameters, errlog: TextIO = sys.stderr
) -> AsyncIterator[
tuple[
MemoryObjectReceiveStream[SessionMessage | Exception],
MemoryObjectSendStream[SessionMessage],
]
]:
"""Drop-in for ``mcp.client.stdio.stdio_client`` with a tolerant reader.
All differences from upstream live in ``stdout_reader``: lines that fail
``JSONRPCMessage.model_validate_json`` are logged at DEBUG and skipped,
never forwarded as exceptions. This is what makes warning-noisy MCPs
quiet from the consumer's point of view.
"""
read_stream_writer, read_stream = anyio.create_memory_object_stream[
SessionMessage | Exception
](0)
write_stream, write_stream_reader = anyio.create_memory_object_stream[
SessionMessage
](0)
try:
command = _get_executable_command(server.command)
env_default = _default_inherited_env()
process = await _create_platform_compatible_process(
command=command,
args=server.args,
env=(
{**env_default, **server.env}
if server.env is not None
else env_default
),
errlog=errlog,
cwd=server.cwd,
)
except OSError:
await read_stream.aclose()
await write_stream.aclose()
await read_stream_writer.aclose()
await write_stream_reader.aclose()
raise
async def stdout_reader() -> None:
assert process.stdout, "Opened process is missing stdout" # noqa: S101
try:
async with read_stream_writer:
buffer = ""
async for chunk in TextReceiveStream(
process.stdout,
encoding=server.encoding,
errors=server.encoding_error_handler,
):
lines = (buffer + chunk).split("\n")
buffer = lines.pop()
for line in lines:
stripped = line.strip()
if not stripped:
continue
try:
message = types.JSONRPCMessage.model_validate_json(
stripped
)
except Exception: # noqa: BLE001 — by design, see module doc
_log.debug(
"lenient stdio: dropped non-JSON line: %r",
stripped[:200],
)
continue
await read_stream_writer.send(SessionMessage(message))
except anyio.ClosedResourceError:
await anyio.lowlevel.checkpoint()
async def stdin_writer() -> None:
assert process.stdin, "Opened process is missing stdin" # noqa: S101
try:
async with write_stream_reader:
async for session_message in write_stream_reader:
payload = session_message.message.model_dump_json(
by_alias=True, exclude_none=True
)
await process.stdin.send(
(payload + "\n").encode(
encoding=server.encoding,
errors=server.encoding_error_handler,
)
)
except anyio.ClosedResourceError:
await anyio.lowlevel.checkpoint()
async with (
anyio.create_task_group() as tg,
process,
):
tg.start_soon(stdout_reader)
tg.start_soon(stdin_writer)
try:
yield read_stream, write_stream
finally:
# MCP spec stdio shutdown: close stdin → wait → SIGTERM → SIGKILL.
if process.stdin:
with contextlib.suppress(Exception):
await process.stdin.aclose()
try:
with anyio.fail_after(PROCESS_TERMINATION_TIMEOUT):
await process.wait()
except TimeoutError:
await _terminate_process_tree(process)
except ProcessLookupError:
pass
await read_stream.aclose()
await write_stream.aclose()
await read_stream_writer.aclose()
await write_stream_reader.aclose()
def _default_inherited_env() -> dict[str, str]:
"""Same env shortlist as ``mcp.client.stdio.get_default_environment``.
Re-exported so we can compose with the user's ``env`` overrides without
forcing a private import in this module's body.
"""
from mcp.client.stdio import get_default_environment
return get_default_environment()
__all__ = ["LenientStdioTransport"]
+27 -4
View File
@@ -23,20 +23,36 @@ class _BaseMcp(BaseModel):
class StdioMcp(_BaseMcp):
"""Subprocess MCP server we spawn and connect to over stdio."""
"""Subprocess MCP server we spawn and connect to over stdio.
``lenient`` switches the upstream stdio reader to a tolerant variant
that silently drops non-JSON-RPC lines from the subprocess's stdout
(``Processing...``-style chatter, banners, dependency-load messages).
The reference ``mcp.client.stdio.stdio_client`` forwards those parse
failures as exceptions, which bleed into Cursor/Cline UIs as warnings
when they connect through us. Default ``False`` keeps the strict
contract — flip it on per-namespace for known-noisy upstreams.
"""
kind: Literal["stdio"] = "stdio"
command: tuple[str, ...]
env: dict[str, str] | None = None
cwd: Path | None = None
lenient: bool = False
class HttpMcp(_BaseMcp):
"""Remote MCP server reached over streamable HTTP."""
"""Remote MCP server reached over streamable HTTP.
``headers`` are forwarded on every request to the upstream MCP — handy
for upstreams that authenticate via custom header rather than a Bearer
token (``auth``).
"""
kind: Literal["http"] = "http"
url: str
auth: str | None = None
headers: dict[str, str] | None = None
class PythonToolMcp(_BaseMcp):
@@ -62,19 +78,26 @@ class McpServer:
command: Iterable[str],
env: dict[str, str] | None = None,
cwd: Path | str | None = None,
lenient: bool = False,
) -> StdioMcp:
return StdioMcp(
name=name,
command=tuple(command),
env=env,
cwd=Path(cwd) if isinstance(cwd, str) else cwd,
lenient=lenient,
)
@classmethod
def http(
cls, *, name: str, url: str, auth: str | None = None
cls,
*,
name: str,
url: str,
auth: str | None = None,
headers: dict[str, str] | None = None,
) -> HttpMcp:
return HttpMcp(name=name, url=url, auth=auth)
return HttpMcp(name=name, url=url, auth=auth, headers=headers)
@classmethod
def python_tool(