Issue: Azure API Management always waits ~180 seconds before responding to MCP requests (even with return-response)

[sebarach]

I am trying to expose a Model Context Protocol (MCP) endpoint through Azure API Management (APIM) using JSON-RPC over HTTP.

Regardless of the logic implemented, every request takes ~3 minutes (≈180 seconds) to return a response, even when the response is fully generated inside APIM policies and should return immediately.

This happens consistently and reproducibly.

Environment

Azure API Management (Consumption / Developer tier)

JSON-RPC 2.0 (MCP protocol)

HTTP POST endpoints

Tested with Postman and Claude Code

Backend APIs are healthy and respond immediately when called directly

What I’m trying to do

Expose an MCP endpoint such as:

POST /mcp
POST /healthcheck/mcp

That handles JSON-RPC methods like:

initialize

tools/list

tools/call

Some of these (e.g. tools/list, healthcheck) are fully handled inside APIM policies and should not call any backend.

Expected behavior

Requests handled entirely in APIM policies should return immediately (<100 ms).

APIM should not wait for backend timeouts if:

A response body is set

Or is used

Actual behavior

Every request takes ~180 seconds before returning

This happens even when:

The response is generated entirely in policy

is explicitly used

No backend is configured or required

After ~3 minutes, APIM finally returns 200 OK with the correct body

This delay happens 100% of the time.

Evidence

The delay is always ~3 minutes (≈180s), strongly suggesting an APIM timeout

Happens even on a brand new test endpoint (/healthcheck/mcp)

Happens even when responding only with tools/list

Happens even when no backend call is needed

Backend APIs respond instantly when called directly

What we already tried
✔ MCP implementation

initialize works

tools/list returns valid MCP response

tools/call logic implemented

✔ APIM policy experiments

Parsing request body with preserveContent: true

Defensive JSON parsing (try/catch)

Setting set-backend-service explicitly

Removing backend entirely

Using to short-circuit the pipeline

Moving logic between inbound / outbound

Separate APIs for MCP and healthcheck

Minimal policies (no routing, no backend, only return-response)

✔ Validation

Same behavior in multiple endpoints

Same behavior with minimal policies

Same behavior regardless of request content

Same behavior even when response body is static

Example request
{
"jsonrpc": "2.0",
"id": 1,
"method": "tools/list"
}

Example policy (simplified)

application/json { "jsonrpc": "2.0", "id": 1, "result": { "ok": true } }

➡️ Still takes ~180 seconds to respond

Why this looks like a platform issue

should immediately terminate the pipeline

There is no backend dependency in these flows

The behavior is deterministic (always ~180s)

It affects multiple endpoints and APIs

It reproduces even with minimal configuration

This makes it very difficult (or impossible) to reliably host MCP / JSON-RPC endpoints on APIM.

Question

Is this a known limitation of Azure API Management?

Is there a required setting to prevent APIM from waiting for backend timeouts when using return-response?

Is APIM officially unsupported for JSON-RPC / MCP-style endpoints?

Any guidance or confirmation would be appreciated.