This page describes the current transitional routing behavior. Its live upstream fan-out and durable-session assumptions are not the Phase 3 target. See ContextForge 2.0 Target Architecture and Roadmap for the proposed ownership boundary and migration.
Backend map keys become public identifiers only for multi-backend virtual hosts without an explicit tool alias:
backend tool "increment" on backend "gateway-one" → "gateway-one-increment"
backend resource "counter" on backend "gateway-one" → "gateway-one-counter"
Single-backend virtual hosts: identifiers pass through unchanged.
Breaking change rule: changing a backend map key changes downstream identifiers for multi-backend virtual hosts. Do not rename without updating merge logic, split logic, and tests.
BackendMCPGateway.tool_name_aliases maps {downstream_alias: upstream_original}. Aliases take precedence over prefix fallback. They are advertised and routed exactly as published (case, dots, underscores preserved).
All four list methods fan out to all connected backends concurrently and merge results:
list_tools / list_resources / list_prompts / list_resource_templates
→ all connected backends → merged sorted output
Failed/unavailable backends are logged and skipped. Single-backend: identifiers unchanged. Multi-backend: prefixed with backend map key.
Calls targeting one object use the inverse rule. The name splitter walks configured backend names and requires a - immediately after the backend name:
gateway-one-increment → backend: gateway-one, tool: increment
gateway-oneincrement → rejected (no - separator)
call_tool resolves explicit alias first, then falls back to single/multi-backend logic.
Methods using the same conditional routing: read_resource, subscribe, unsubscribe, get_prompt, complete.
The gateway wraps per-backend cursors inside its own opaque token (JSON, treated as opaque by MCP clients). First request: all backends queried. Resume: cursor decoded, exhausted backends skipped. New cursor emitted when any backend has more pages.
Known limitation: if backend set changes between pages, removed backend's cursor is silently dropped.
Backend RMCP services are stored in BackendTransports keyed by:
principal (claims.sub) + backend_name (map key) + downstream_session_id
This is local process state only. Implications:
- After
initialize, later requests must reach the same process. - Sticky routing required for load-balanced deployments.
- Gateway restart → all sessions lost → clients must re-run
initialize. - Multi-runtime mode (
--single-runtime false): each runtime thread has its ownBackendTransportswith no cross-thread affinity.
Exception: call_tool uses per-request backend lifecycle. Each tool call creates a fresh backend connection, executes the call with plugin hooks, then closes the connection. This bypasses BackendTransports entirely and does not require session affinity for tool calls specifically (though other MCP methods still do).
sequenceDiagram
participant C as MCP Client
participant GW as Gateway (RMCP)
participant BT as BackendTransports<br/>(local process state)
participant LU as LocalUserSessionStore<br/>(LRU 50k / 1h)
participant BA as Backend A
participant BB as Backend B
C->>GW: POST initialize (Mcp-Session-Id: S)
GW->>BA: initialize (concurrent)
GW->>BB: initialize (concurrent)
BA-->>GW: InitializeResult
BB-->>GW: InitializeResult
GW->>BT: store RunningService keyed by sub+backend+S
GW->>LU: store session entry for sub+S
GW-->>C: merged InitializeResult
C->>GW: POST call_tool (Mcp-Session-Id: S)
GW->>BT: lookup sub+backend+S → Arc<RunningService>
BT-->>GW: RunningService handle
GW->>BA: call_tool (routed by name prefix)
BA-->>GW: ToolResult
GW-->>C: ToolResult
C->>GW: DELETE (Mcp-Session-Id: S)
GW->>GW: RMCP handles DELETE
GW->>LU: remove sub+S entry
GW->>BT: remove all sub+*+S entries
GW-->>C: 200 OK
On initialize, the gateway builds one downstream InitializeResult — not a passthrough of any one backend. The source of truth is each backend's InitializeResult; the gateway reads peer_info().capabilities from each running service and stores them with the backend transport state.
The merge rule (gateway-aware, not a raw union):
- Enable a top-level capability when ≥1 backend supports it and the gateway has a routing story for it.
resources.subscribepreserved if any backend advertises it (the gateway routes subscribe/unsubscribe and forwards resource-update notifications).listChangednot yet advertised (gateway doesn't emit downstream list-changed notifications when upstream lists change).- Single-backend passthrough is not a stable contract (
HashMapiteration order). - If no backend reports supported capabilities, returns
ServerCapabilities::default().
Do not initialize the downstream capability from just one backend entry — the gateway fronts multiple backends, HashMap iteration is non-deterministic, and list methods already merge across all backends.
DELETE with Mcp-session-id:
→ RMCP handles request
→ on success: remove LocalUserSessionStore entry + BackendTransports entries for principal+session
If RMCP rejects the delete, local state is untouched.
| Method | Group | Behavior |
|---|---|---|
initialize |
Session | Concurrent fanout to all backends; failure of one backend is non-fatal (stored with no service). Returns merged capability set. Requires DownstreamSessionId, UserConfig, VirtualHostId, ContextForgeClaims. |
list_tools |
List | Fan-out all connected backends → merged sorted result. Cursor-based pagination across backends. |
list_resources |
List | Same as list_tools. |
list_prompts |
List | Same as list_tools. |
list_resource_templates |
List | Same — both name and URI template get prefixed for multi-backend. |
call_tool |
Targeted | Per-request backend lifecycle: creates fresh connection via connect_backend_for_request, runs pre-hook, executes call, runs post-hook, closes connection. Resolves alias → single/multi-backend fallback. Forwards downstream cancellation to backend. Tracks backend progress tokens: RMCP assigns a new token per backend request; the gateway maps each backend token to the downstream token. Request enqueue and mapping publication are serialized against progress lookup so an immediate backend notification cannot overtake registration. When the notification matches an in-flight token, the gateway restores the downstream token and forwards it to the client. Does not use session-backed BackendTransports. |
read_resource |
Targeted | Single-backend: URI unchanged. Multi-backend: strips prefix. |
subscribe / unsubscribe |
Targeted | Same resource-URI routing; forwards/stops resource-update notifications. |
get_prompt |
Targeted | Single-backend: name unchanged. Multi-backend: strips prefix. Runs pre/post prompt hooks around the backend call: the pre hook may rewrite arguments or deny, the post hook may rewrite or reject the rendered messages. |
complete |
Targeted | Routes on prompt name or resource URI inside ref. |
ping |
Local | Returns success; no backend fanout. |
DELETE |
Session | RMCP handles first; on success session_id_layer removes local session + backend transports. |