McpToolset connection failure (HTTP 401/403/5xx) causes ASGI crash instead of graceful error

🔴 Required Information

Describe the Bug:
When an MCP server returns HTTP 401 Unauthorized (e.g., due to an expired/invalid JWT token), the entire ASGI application crashes instead of returning a proper error to the caller. The HTTP error is not propagated, causing the SSE response stream to terminate abruptly.

Steps to Reproduce:

Configure McpToolset with a header_provider that returns a JWT token
Make a request where the JWT is invalid/expired (or MCP server rejects it)
The MCP server returns HTTP 401
Observe app crash with ASGI callable returned without completing response

from google.adk.tools.mcp_tool import McpToolset, StreamableHTTPConnectionParams

def _mcp_header_provider(ctx: ReadonlyContext) -> dict[str, str]:
    jwt = ctx.state.get("jwtToken")
    if not jwt:
        return {}
    return {"Authorization": f"Bearer {jwt}"}

toolset = McpToolset(
    connection_params=StreamableHTTPConnectionParams(
        url="https://example.com/mcp",
    ),
    header_provider=_mcp_header_provider,
    tool_filter=["some_tool"],
)

Expected Behavior:
The 401 error should be caught and propagated as a ConnectionError or similar, allowing the application to handle it gracefully (e.g., return an error message to the user, refresh token, etc.).

Observed Behavior:
The application crashes with:

ERROR - uvicorn.error - ASGI callable returned without completing response.

Environment Details:

ADK Library Version: 1.20.0
Desktop OS: Linux
Python Version: 3.13.8
MCP library version: 1.25.0

Model Information:

Are you using LiteLLM: No
Which model is being used: N/A (crash happens before LLM interaction)

🟡 Optional Information

Logs:

2026-01-27 16:18:49,502 DEBUG - mcp.client.streamable_http - Connecting to StreamableHTTP endpoint: https://example.com/mcp
2026-01-27 16:18:49,502 DEBUG - mcp.client.streamable_http - Sending client message: root=JSONRPCRequest(method='initialize', ...)
2026-01-27 16:18:49,583 DEBUG - httpcore.http11 - receive_response_headers.complete return_value=(b'HTTP/1.1', 401, b'', [..., (b'WWW-Authenticate', b'Basic realm="Realm"'), ...])
2026-01-27 16:18:49,584 DEBUG - httpcore.http11 - response_closed.complete
2026-01-27 16:18:49,585 ERROR - uvicorn.error - ASGI callable returned without completing response.

Additional Context:
The issue may originate in the MCP Python SDK (streamable_http.py) where HTTP errors like 401 are not sent through the response stream, unlike 404 which is handled explicitly. This might warrant an upstream issue to https://github.com/modelcontextprotocol/python-sdk as well.

How often has this issue occurred?:

Always (100%) — whenever MCP server returns a non-2xx/non-404 status code

Seeing this same behavior when a remote mcp server becomes unavailable. Right now it seems like the adk assumes that the mcp server will always be available and will return a 2xx response. Would be nice to be able to handle this from the adk

@simone-viozzi

@simone-viozzi Thanks for reporting!
Validate the HTTP response status before entering the SSE loop. If the server returns a non-2xx status, raise an exception immediately so the ASGI response is not left incomplete.

async with client.stream("POST", url, headers=headers, json=payload) as response:
    if not (200 <= response.status_code < 300):
        body = await response.aread()
        raise ConnectionError(
            f"MCP server returned HTTP {response.status_code}: "
            f"{body.decode(errors='ignore')}"
        )

    async for line in response.aiter_lines():
        if line.strip():
            yield line

This prevents stream interruption on 401/403/500 responses and avoids the ASGI crash
Thank you.

Validate the HTTP response status before entering the SSE loop. If the server returns a non-2xx status, raise an exception immediately so the ASGI response is not left incomplete.

Sorry, where should I validate the request?

Do I need to override StreamableHTTPConnectionParams?

I'm using the ADK interface to debug.

@simone-viozzi

Hi @simone-viozzi,

could you try update adk to latest version(1.24.1) and reproduce your issue? Recent changes with handling MCP client at #4025 may resolve your issue.

You may also check the comment here for more details.

@simone-viozzi

@simone-viozzi This was fixed in v1.23.0 via PR #4025 — SessionContext now wraps the MCP session lifecycle in a dedicated task, so HTTP errors during initialization propagate as ConnectionError instead of crashing the ASGI app.

Upgrade to v1.24.1 (pip install google-adk==1.24.1) and wrap your get_tools() call in try/except ConnectionError to handle auth failures gracefully.

The underlying MCP SDK still doesn't validate HTTP status before the SSE loop (streamable_http.py), but the ADK-side error propagation now prevents the crash you're seeing.

@challenger71498

Confirmed fixed in v1.23.0 (PR #4025) — tested on v1.24.1 with mock servers returning 401/403/500, all now raise ConnectionError instead of crashing the ASGI app.

The SessionContext wrapper at session_context.py:189 catches the MCP SDK error and propagates it through mcp_toolset.py:286-287.

@challenger71498 also pointed to the same fix.

Closing this issue for now as completed.

Navigation Menu

Repository navigation

McpToolset connection failure (HTTP 401/403/5xx) causes ASGI crash instead of graceful error #4300

🔴 Required Information

🟡 Optional Information

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

McpToolset connection failure (HTTP 401/403/5xx) causes ASGI crash instead of graceful error #4300

Description

🔴 Required Information

🟡 Optional Information

Activity

jfox1-02 commented on Feb 5, 2026

klateefa commented on Feb 6, 2026

simone-viozzi commented on Feb 6, 2026

challenger71498 commented on Feb 7, 2026

surajksharma07 commented on Feb 8, 2026

surajksharma07 commented on Feb 9, 2026

Metadata

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

Issue actions