Unpack settings in FastMCP #1198

Kludex · 2025-07-25T09:59:35Z

No description provided.

Kludex · 2025-07-25T10:00:05Z

src/mcp/server/fastmcp/server.py

+        debug: bool = False,
+        log_level: Literal["DEBUG", "INFO", "WARNING", "ERROR", "CRITICAL"] = "INFO",
+        host: str = "127.0.0.1",
+        port: int = 8000,
+        mount_path: str = "/",
+        sse_path: str = "/sse",
+        message_path: str = "/messages/",
+        streamable_http_path: str = "/mcp",
+        json_response: bool = False,
+        stateless_http: bool = False,
+        warn_on_duplicate_resources: bool = True,
+        warn_on_duplicate_tools: bool = True,
+        warn_on_duplicate_prompts: bool = True,
+        dependencies: Collection[str] = (),
+        lifespan: Callable[[FastMCP[LifespanResultT]], AbstractAsyncContextManager[LifespanResultT]] | None = None,
+        auth: AuthSettings | None = None,
+        transport_security: TransportSecuritySettings | None = None,


This is the main change. The rest is just to comply with those types.

dsp-ant · 2025-07-25T10:12:33Z

src/mcp/server/fastmcp/server.py

+        port: int = 8000,
+        mount_path: str = "/",
+        sse_path: str = "/sse",
+        message_path: str = "/messages/",


Suggested change

message_path: str = "/messages/",

message_path: str = "/messages",

Hard to trust AI nowadays. 👀

Oh, actually, the AI didn't make a mistake.

The default on Settings.message_path was added with the /.

yea i know :(, is that correct tho?

Given that it doesn't follow the pattern, no. But also there's some redirect issues that people may be facing.

But should I really change it in this PR? It may affect people.

dsp-ant · 2025-07-25T10:18:00Z

src/mcp/server/fastmcp/server.py

+        log_level: Literal["DEBUG", "INFO", "WARNING", "ERROR", "CRITICAL"] = "INFO",
+        host: str = "127.0.0.1",
+        port: int = 8000,
+        mount_path: str = "/",
+        sse_path: str = "/sse",
+        message_path: str = "/messages/",
+        streamable_http_path: str = "/mcp",
+        json_response: bool = False,
+        stateless_http: bool = False,
+        warn_on_duplicate_resources: bool = True,
+        warn_on_duplicate_tools: bool = True,
+        warn_on_duplicate_prompts: bool = True,
+        dependencies: Collection[str] = (),
+        lifespan: Callable[[FastMCP[LifespanResultT]], AbstractAsyncContextManager[LifespanResultT]] | None = None,
+        auth: AuthSettings | None = None,
+        transport_security: TransportSecuritySettings | None = None,


My main worry is that we now keep the defaults in sync between Settings and this. I think ideally, but sadly a BC break would be to just accept settings: Settings = Settings() here. Since I think that would be backwards incompatible. For now, since Settings is only used internally to FastMCP should we remove the defaults from the Settings class, such that we only provide it here?

Yep, we should remove the defaults from Settings.

dsp-ant

LGTM

GZGavinZhao · 2025-07-31T21:08:14Z

This PR likely broke a lot of uvx my-mcp-server workflows by removing **settings as an argument to FastMCP.__init__:

Traceback (most recent call last):
  File "/home/gavinzhao/.cache/uv/archive-v0/Sd9MpTigriP_4YpcGN7U3/bin/lean-lsp-mcp", line 6, in <module>
    from lean_lsp_mcp import main
  File "/home/gavinzhao/.cache/uv/archive-v0/Sd9MpTigriP_4YpcGN7U3/lib64/python3.13/site-packages/lean_lsp_mcp/__init__.py", line 3, in <module>
    from lean_lsp_mcp.server import mcp
  File "/home/gavinzhao/.cache/uv/archive-v0/Sd9MpTigriP_4YpcGN7U3/lib64/python3.13/site-packages/lean_lsp_mcp/server.py", line 98, in <module>
    mcp = FastMCP(**mcp_kwargs)
TypeError: FastMCP.__init__() got an unexpected keyword argument 'env_vars'

I understand that it's kind of awkward to have a backwards compatibility guarantee for a fast-moving package like mcp, but perhaps it'd be a good idea to at least document this change? Especially since this change is made in a patch release?

MischaPanch · 2025-07-31T21:13:20Z

This is a breaking change and you only updated the patch version of the package. It broke our code:
oraios/serena#381

You could fix this by rolling this back, urgently releasing 1.12.4, and then doing it in a backwards compatible way or make a new major version.

For reference of what breaks: this code now causes a pydantic validation error

mcp_settings: Settings = Settings(lifespan=self.server_lifespan, host=host, port=port)
mcp = FastMCP(**mcp_settings.model_dump(), instructions="...")

MischaPanch · 2025-07-31T21:18:54Z

Fun coincidence: I had raised the problem with the constructor a while ago and in the issue had proposed a backwards compatible way to handle this
#305

GZGavinZhao · 2025-07-31T21:21:38Z

For uvx users, a temporary fix for this is to have mcp==1.12.2 in a requirements.txt file and add --with-requirements=requirements.txt as the first arugment to uvx.

Actually, you can also add the environment variable UV_CONSTRAINT=<path-to-requirements.txt-above> to whatever MCP configuration you're using, e.g.:

{
  "mcpServers": {
    "lean-lsp": {
      "type": "stdio",
      "command": "uvx",
      "args": [
        "lean-lsp-mcp"
      ],
      "env": {
        "UV_CONSTRAINT": "requirements.txt"
      }
    }
  }
}

mzyil · 2025-08-02T21:34:41Z

please add a bot to your repo that verifies backwards compat. a lot of people depend on your code and you are past alpha.

Kludex · 2025-08-03T06:34:01Z

It seems there were two different issues here:

Use of env_vars (i.e. passing unused parameter to the constructor)
Instantiating Settings by oneself, which I guess would be to have a false sense of type safety?

I didn't expect people to be instantiating the Settings themselves.

I guess the way to go here now is to:

Ignore the messages above - this change seems reasonable.
Add a kwargs back with a deprecation warning in it saying that those parameters are not used, and add the default values for settings back - which I think we should deprecate the Settings class itself.

MischaPanch · 2025-08-03T08:18:01Z

Settings is a public class and it was the only way to have the kwargs names and types analyzed by the IDE - how was the sense of type safety false?

But it seems like indeed few people were using it, otherwise this would have gotten more attention

Unpack settings in FastMCP

a6f413d

Kludex requested a review from a team as a code owner July 25, 2025 09:59

Kludex requested a review from ochafik July 25, 2025 09:59

Kludex commented Jul 25, 2025

View reviewed changes

Kludex requested a review from ihrpr July 25, 2025 10:00

Kludex mentioned this pull request Jul 25, 2025

Set LifespanResultT default=Any to prevent reportMissingTypeArgument #1139

Closed

9 tasks

dsp-ant reviewed Jul 25, 2025

View reviewed changes

Drop defaults from Settings

2222463

dsp-ant approved these changes Jul 25, 2025

View reviewed changes

Kludex merged commit 959d4e3 into main Jul 26, 2025
13 checks passed

Kludex deleted the unpack-settings-fastmcp branch July 26, 2025 08:37

gspencergoog pushed a commit to gspencergoog/mcp-python-sdk that referenced this pull request Jul 29, 2025

Unpack settings in FastMCP (modelcontextprotocol#1198)

7b3a96a

LukePeltier mentioned this pull request Jul 31, 2025

MCP Server Pydantic validation error after recent merge oraios/serena#381

Closed

GZGavinZhao mentioned this pull request Jul 31, 2025

mcp v1.12.3 broke lean-lsp-mcp oOo0oOo/lean-lsp-mcp#19

Closed

neelmitra mentioned this pull request Aug 1, 2025

DynamoDB: FastMCP.__init__() got an unexpected keyword argument 'version' awslabs/mcp#996

Open

1 task

liukatkat mentioned this pull request Aug 1, 2025

feat: core RPC scanning semgrep/mcp#87

Merged

LarissaSmith mentioned this pull request Aug 4, 2025

Claude AI desktop no longer connecting to Blender MCP server? ahujasid/blender-mcp#144

Open

	message_path: str = "/messages/",
	message_path: str = "/messages",

Unpack settings in FastMCP #1198

Unpack settings in FastMCP #1198

Conversation

Kludex commented Jul 25, 2025

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

dsp-ant left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

GZGavinZhao commented Jul 31, 2025

Uh oh!

MischaPanch commented Jul 31, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

MischaPanch commented Jul 31, 2025

Uh oh!

GZGavinZhao commented Jul 31, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

mzyil commented Aug 2, 2025

Uh oh!

Kludex commented Aug 3, 2025

Uh oh!

MischaPanch commented Aug 3, 2025

Uh oh!

Uh oh!

MischaPanch commented Jul 31, 2025 •

edited

Loading

GZGavinZhao commented Jul 31, 2025 •

edited

Loading