Skip to content

MCP servers fail to connect after mcp SDK 2.0.0: Tool.inputSchema renamed to input_schema #345

Description

@Joi

Summary

mcp 2.0.0 was published to PyPI on 2026-07-28. amplifier-module-tool-mcp declares an unbounded mcp>=1.0.0, so any install that resolves dependencies after that date picks up 2.0.0 and every configured MCP server stops connecting.

Filing here because issues are disabled on microsoft/amplifier-module-tool-mcp. Fix PR: microsoft/amplifier-module-tool-mcp#10

Symptom

Failed to connect to MCP server 'qmd': unhandled errors in a TaskGroup (1 sub-exception)
Failed to start MCP server 'qmd': MCP server connection failed: unhandled errors in a TaskGroup (1 sub-exception)
Failed to connect to MCP server 'qmd-pim': unhandled errors in a TaskGroup (1 sub-exception)
Failed to connect to Streamable HTTP MCP server 'senzing': 'Tool' object has no attribute 'inputSchema'
Failed to connect to MCP server 'msgvault': unhandled errors in a TaskGroup (1 sub-exception)

The two different error strings are one root cause. _discover_capabilities() runs inside the SDK's anyio TaskGroup on stdio, so the AttributeError is swallowed and resurfaces as the opaque unhandled errors in a TaskGroup (1 sub-exception). Only the streamable-HTTP transport reports it plainly.

This is worth calling out for anyone else debugging it: the server-side logs are clean. ~/.amplifier/logs/mcp-servers/msgvault.log showed a normal startup and exit outcome=ok at the exact failure timestamp — the server launched fine and exited cleanly; the client crashed while listing tools. That makes it look like a transport or timeout problem when it is a Python attribute rename.

Root cause

mcp 2.0.0 renames the mcp.types model fields from camelCase to snake_case. The JSON wire names are preserved as pydantic aliases, so the protocol is unchanged — only Python attribute access changed:

Wire field mcp 1.x attribute mcp >= 2.0 attribute
inputSchema .inputSchema .input_schema
mimeType .mimeType .mime_type

Two affected call sites in amplifier-module-tool-mcp v0.2.2, with different failure modes:

  • Tool.inputSchema (client.py:321, streamable_http_client.py:278) — hard AttributeError. Kills capability discovery, so connection fails for every server, on both transports.
  • Resource.mimeType (client.py:334, streamable_http_client.py:291) — silent, not a crash. Both sites were already hasattr(resource, "mimeType")-guarded, so on 2.x the guard falls through and every discovered resource reports mime_type: None.

All other mcp API surface this module depends on is unaffected — I verified ClientSession, StdioServerParameters, mcp.client.stdio.stdio_client, mcp.client.streamable_http.streamable_http_client, and mcp.shared._httpx_utils.create_mcp_http_client all still import on 2.0.0, and that CallToolResult.content / ReadResourceResult.contents / GetPromptResult.messages / PromptArgument.{name,description,required} keep their names.

Reproduction

uv venv /tmp/repro -p 3.12
uv pip install --python /tmp/repro/bin/python "amplifier-core==1.6.0" "mcp==2.0.0"
uv pip install --python /tmp/repro/bin/python -e /path/to/amplifier-module-tool-mcp --no-deps
# then connect MCPClient to any stdio MCP server and call connect()

Downgrading to mcp==1.29.0 restores service, which is a usable interim workaround for anyone blocked right now:

uv pip install --python ~/.local/share/uv/tools/amplifier/bin/python "mcp==1.29.0"

That pin is fragile, though — the next amplifier update / uv tool upgrade amplifier can re-resolve and pull 2.0.0 again.

Fix

PR against amplifier-module-tool-mcp adds a small sdk_compat.py with get_input_schema() / get_mime_type() accessors that read whichever attribute the installed SDK exposes, so the module works unchanged on both majors. Verified: full suite passes on mcp==1.29.0 and mcp==2.0.0 (64 each), and end-to-end capability discovery against a real stdio MCP server succeeds on both.

Impact

Anyone who installs or updates Amplifier with MCP servers configured after 2026-07-28. It is not machine-specific and not config-specific — the whole fleet is one dependency resolve away from it.

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions