McpServer.layerStdio: two stdio servers in one layer graph share one stdio protocol, so the second never reads its stdin

[spencerbeggs]

This may well be intended: a process has one stdin, so one stdio server per process is a natural model. We're reporting it because we hit it in tests, where two servers each get their own Stdio and the result surprised us.

What we observed

We merged two McpServer.layerStdio servers into one layer graph with Layer.mergeAll, each with its own Stdio. Only the first server answers. The second never reads its own stdin. If the same two servers are built separately, each with its own Layer.build, both answer.

Environment: effect@4.0.0-rc.117 (tag effect@4.0.0-rc.117, 14a3f14), on Node 26.10.0 and macOS. From reading the source, layerProtocolStdio is still a module-level constant on main at 3788b63 (packages/effect/src/rpc/RpcServer.ts:1436-1440).

Reproduction

import { Effect, Layer, Sink, Stdio, Stream } from "effect"
import { McpProtocol, McpServer } from "effect/unstable/ai"

const initialize = (id: number) =>
  JSON.stringify({
    jsonrpc: "2.0",
    id,
    method: "initialize",
    params: { protocolVersion: "2025-11-25", capabilities: {}, clientInfo: { name: "repro", version: "0" } }
  }) + "\n"

const makeStdio = (name: string, id: number, out: Array<string>) =>
  Stdio.layerTest({
    stdin: Stream.make(new TextEncoder().encode(initialize(id))).pipe(Stream.concat(Stream.never)),
    stdout: () =>
      Sink.forEach((chunk: string | Uint8Array) =>
        Effect.sync(() =>
          out.push(`${name} stdout: ${typeof chunk === "string" ? chunk : new TextDecoder().decode(chunk)}`.trim())
        )
      )
  })

const server = (name: string) =>
  McpServer.layerStdio({ name, version: "0.0.0", protocols: [McpProtocol.v2025_11_25] })

const report = (label: string, outA: Array<string>, outB: Array<string>) =>
  Effect.sync(() => {
    console.log(`--- ${label}`)
    console.log(outA.length ? outA.join("\n").slice(0, 150) : "A stdout: (nothing)")
    console.log(outB.length ? outB.join("\n").slice(0, 150) : "B stdout: (nothing)")
  })

const outA: Array<string> = []
const outB: Array<string> = []

const program = process.argv[2] === "separate"
  ? Effect.gen(function*() {
    yield* Layer.build(server("A").pipe(Layer.provide(makeStdio("A", 1, outA))))
    yield* Layer.build(server("B").pipe(Layer.provide(makeStdio("B", 2, outB))))
    yield* Effect.sleep("300 millis")
    yield* report("each server in its own Layer.build", outA, outB)
  })
  : Effect.gen(function*() {
    yield* Layer.build(Layer.mergeAll(
      server("A").pipe(Layer.provide(makeStdio("A", 1, outA))),
      server("B").pipe(Layer.provide(makeStdio("B", 2, outB)))
    ))
    yield* Effect.sleep("300 millis")
    yield* report("both servers in one Layer.mergeAll", outA, outB)
  })

Effect.runFork(Effect.scoped(program))

We ran node repro.ts separate and node repro.ts merged, and captured the output live:

--- each server in its own Layer.build
A stdout: {"jsonrpc":"2.0","id":1,"result":{"protocolVersion":"2025-11-25","capabilities":{"logging":{},"completions":{}},"serverInfo":{"name":"A","ve
B stdout: {"jsonrpc":"2.0","id":2,"result":{"protocolVersion":"2025-11-25","capabilities":{"logging":{},"completions":{}},"serverInfo":{"name":"B","ve
--- both servers in one Layer.mergeAll
A stdout: {"jsonrpc":"2.0","id":1,"result":{"protocolVersion":"2025-11-25","capabilities":{"logging":{},"completions":{}},"serverInfo":{"name":"A","ve
B stdout: (nothing)

Our reading

No spec is involved here. This is a question about layer design. MCP's stdio transport has the client launch "the MCP server as a subprocess", so one server per process is the normal production shape. We hit this only in tests and tooling that host several servers in one process.

Where it happens

• RpcServer.ts:1416-1420: layerProtocolStdio is a module-level Layer.effect(Protocol, makeProtocolStdio).
• McpServer.ts:1439-1447: every layerStdio(...) call provides that same constant. Layers are memoized by reference within one build, so as we read it, both servers resolve Protocol to a single instance built over whichever Stdio was reached first. That fits what we see: B's initialize (id 2) is answered on neither stdout.

How we work around it

We don't have a code fix. In @effected/mcp we document the limitation: give each server a fresh layer memo map, through its own ManagedRuntime, Effect.provide(layer, { local: true }), or its own process.

Is this intended?

If one stdio server per memo map is the intended model, a sentence in the layerStdio / layerProtocolStdio docs would have saved us some time. If it isn't intended, one possible direction is for layerStdio to build its protocol layer per call rather than sharing the module constant. We haven't worked through how that would interact with the shared McpServer.layer.

[IMax153]

@spencerbeggs - I've opened #8506 but I'd like to get @tim-smart 's eyes on it because I'm not sure it's really the direction we want to go. We have many such instances where we provide a Layer "bundle" for convenience, but generally if you want to avoid memoization of certain layers we would recommend building the bundle yourself and using Layer.fresh where you want fresh copies.

[IMax153]

For example, to fix the issue you are having in tests where you want to run two separate servers, you could employ Layer.fresh for the whole layer:

Layer.mergeAll(
  Layer.fresh(server("A").pipe(Layer.provide(makeStdio("A", 1, outA)))),
  Layer.fresh(server("B").pipe(Layer.provide(makeStdio("B", 2, outB))))
)

[spencerbeggs]

@IMax153 Thanks. That works: wrapping each whole bundle in Layer.fresh gives every server its own stdio protocol. We ran the repro on 4.0.0-rc.117 with a toolkit added to each server, because that exposed a trap worth knowing about.

McpServer.toolkit provides McpServer.layer, which is a module-level constant, the same way layerProtocolStdio is. So the tool registry is shared through the memo map just like the protocol, and where the Layer.fresh boundary sits matters:

Shape (one tool per server; initialize, then tools/list)	Result
Both servers merged, no Layer.fresh	only A answers, and it lists both tools
Layer.fresh around each whole bundle (toolkit + layerStdio + Stdio)	each answers and lists only its own tool
Layer.fresh around toolkit + layerStdio, Stdio provided outside	same, correct
Layer.fresh around layerStdio only, toolkit outside	each answers, but lists no tools, even with a single server

In the last row the toolkit registers into the outer memo map's McpServer.layer, and the fresh server builds a second, empty registry. Without tools that shape looks fixed, which is how we first misread it.

For nesting, Layer.build/Effect.provide of a second server under the first one's Effect.provide reuses the first protocol. Layer.fresh around the whole inner bundle, or Effect.provide(layer, { local: true }), separates them.

If #8506 lands, a fresh protocol per layerStdio call fixes the protocol half. The shared registry would still make the Layer.fresh boundary matter for anyone composing toolkits by hand, so a line about "keep the toolkit inside the fresh boundary" in the layerStdio / McpServer.toolkit docs might save the next person the detour. Either way, the recipe you gave is what we now teach.