The exported era classifier. classify_inbound_request(body, headers=...) from mcp.shared.inbound is the body-primary test for "is this a 2026-era request?"; wrap it as classify_era() to route eras to different backends in your own ASGI/ingress layer. Unlike most SDKs, the Python SDK's built-in streamable_http_app() already serves sessionful 2025 alongside stateless 2026 on one /mcp route — so the predicate is for when you need different arms (per-era auth, separate ports, an existing v1 deployment to keep), not to make dual-era work at all.

Also shown: the CORS recipe (methods, request headers, and expose_headers) browser-based MCP clients need.

Run it

# HTTP only — the predicate is an HTTP-transport concern. The client
# self-hosts the app on a free port, runs, then tears it down.
uv run python -m stories.legacy_routing.client --http
# same, against the lowlevel-API server variant
uv run python -m stories.legacy_routing.client --http --server server_lowlevel

# against a server you run yourself (real uvicorn on :8000)
uv run python -m stories.legacy_routing.server --port 8000 &
SERVER_PID=$!
uv run python -m stories.legacy_routing.client --http http://127.0.0.1:8000/mcp
kill "$SERVER_PID"

What to look at

client.py — two visible connections to the SAME /mcp endpoint from one targets() factory: Client(targets(), mode=mode) (default "auto" → server/discover → the modern arm) and Client(targets(), mode="legacy") (the initialize handshake → the legacy arm). Each asserts which_arm reports the era the built-in router actually dispatched to. The era decision is one explicit mode= argument at construction.
client.py — the predicate then shown directly against a modern body, a legacy body, and a malformed-modern body. The runnable build_app() uses the SDK's built-in router; the predicate itself is exercised as a pure function — see the user-land composition recipe below for wiring it into your own ingress.
server.py classify_era — the tri-state wrapper. InboundModernRoute → "modern"; rung-1 INVALID_PARAMS (no envelope keys) → "legacy"; any other InboundLadderRejection is a malformed-modern request to reject, not route to legacy. When headers are supplied, both Mcp-Protocol-Version and Mcp-Method must mirror the body — a disagreement (or an unsupported version) is what produces that third arm; client.py shows both.
server.py build_app — streamable_http_app() + CORSMiddleware. The which_arm tool reads ctx.request_context.protocol_version to prove which path the built-in router took.
server_lowlevel.py — the CORS recipe re-used from server.py (the MCP_* header and method constants); build_app wires lowlevel.Server instead of MCPServer and reads ctx.protocol_version directly. The predicate is tier-agnostic, so classify_era lives only in server.py.

User-land composition (when you need different backends)

There is no legacy="reject" flag yet. To route eras to different handlers, buffer the body, classify, replay:

async def mcp_endpoint(scope, receive, send):
    body, replay = await buffer_body(receive)          # your ASGI helper
    headers = {k.decode("ascii").lower(): v.decode("latin-1") for k, v in scope["headers"]}
    match classify_era(json.loads(body or b"{}"), headers):
        case "legacy":
            await my_existing_v1_manager.handle_request(scope, replay, send)
        case "modern":
            await modern_manager.handle_request(scope, replay, send)
        case rejection:
            await send_jsonrpc_error(send, rejection)  # map via ERROR_CODE_HTTP_STATUS

Non-POST verbs (GET standalone-SSE, DELETE session termination) are sessionful-2025-only — route them straight to the legacy arm.

Two ports instead of one

Run two uvicorn processes from the same build_app() on different ports and put classify_era() (or a header check) in your ingress. Useful when the two eras need different auth, rate limits, or scaling.

Caveats

The SDK's built-in routing is currently header-only — a 2026 client that omits MCP-Protocol-Version is mis-routed to legacy. classify_inbound_request() is body-primary and is what the built-in moves to in a later release; user-land routing with the predicate is already correct today.
ctx.request_context.protocol_version is the interim 2-hop reach; a later release will shorten it.
DNS-rebinding protection is on by default; the harness disables it (NO_DNS_REBIND) because the in-process httpx client sends no Origin. Drop the kwarg for a real deployment.
mcp.shared.inbound is a deep import path — a shorter re-export is planned before beta.

README.md

legacy-routing

Run it

What to look at

User-land composition (when you need different backends)

Two ports instead of one

Caveats

Spec

See also