Added MCP elicitation support for tool injection in mcp-run-python #2258

yamanahlawat · 2025-07-19T17:15:32Z

Closes Tool Injection for mcp-run-python: Enable Python code to call back to agent tools with request/response flow #2037
Closes Support MCP elicitations both in client and server #2330

yamanahlawat · 2025-07-20T15:22:32Z

DouweM · 2025-07-22T00:15:40Z

@yamanahlawat Thank you! This is a great feature, but we'll want to think a bit about where certain code lives. For example, the 2 callbacks methods are too MCP/mcp-run-python-specific to live on Agent.

Can you try to move those to MCPServer? Or, perhaps better, to separate helper functions that a user can explicitly choose to add to their MCPServer if they want to support those specific elicitations?

Those functions then wouldn't have access to self._function_toolset anymore, but I think it'd be better to have to explicitly pass the MCPServer (elicitation callback handler) a toolset to use, and for the case where you want it to have access to the entire agent toolset, we could have a method like agent.set_mcp_sampling_model().

yamanahlawat · 2025-07-22T08:47:24Z

@DouweM Thanks. make sense about the placement. i was initially skeptical about adding these methods directly to the Agent class.

i'll:

move the callback creation logic to separate helper functions in the MCP module.
add a method like set_mcp_sampling_model() (ex: set_mcp_elicitation_toolset) on Agent that applies the callbacks to MCP servers.
remove the auto-magic setup and make it user configured.
update the callbacks to work with explicitly passed toolsets instead of accessing self._function_toolset

and if set_mcp_elicitation_toolset() is called without args, it will default to using the agent's complete toolset.

…lexity

yamanahlawat · 2025-07-25T18:42:47Z

@DouweM Please review.

DouweM

@yamanahlawat Thanks Yaman! This looks a lot better already but I think it'll take a few more cycles until this is ready for merge -- check out my feedback. I also suggest pulling out elicitation_callback into a new PR so we can merge that one quickly and others can start using it, while we work on the rest.

DouweM · 2025-07-29T17:23:46Z

pydantic_ai_slim/pydantic_ai/mcp.py

+                pass
+
+            if available_tools:
+                tool_arguments = {**tool_arguments, 'tools': available_tools}


Can we pass this through metadata instead?

Actually, what if we make mcp-run-python use elicitiation to request the available tools, so that we only need to set an elicitation_callback and not also a process_tool_call?

DouweM · 2025-07-29T17:24:28Z

pydantic_ai_slim/pydantic_ai/mcp.py

@@ -66,6 +235,8 @@ class MCPServer(AbstractToolset[Any], ABC):
    allow_sampling: bool = True
    max_retries: int = 1
    sampling_model: models.Model | None = None
+    allow_elicitation: bool = True
+    elicitation_callback: ElicitationFnT | None = None


Please pull this new elicitation_callback property out into a new PR so we can merge that pretty quickly and close #2330?

DouweM · 2025-07-29T17:25:03Z

pydantic_ai_slim/pydantic_ai/mcp.py

+    )
+
+
+def create_tool_elicitation_callback(toolset: AbstractToolset[Any]) -> ElicitationFnT:


Please move the mcp-run-python specific helper functions to pydantic_ai.ext.mcp_run_python so that this file and its elicitation stuff can stay generic.

DouweM · 2025-07-29T17:30:26Z

pydantic_ai_slim/pydantic_ai/mcp.py

+        self.wrapped = wrapped
+
+        # If it's a CombinedToolset, filter out mcp-run-python toolsets before any get_tools() calls
+        if isinstance(wrapped, CombinedToolset):


This won't work if there are multiple levels of nested toolsets, unfortunately.

Could we require users to specify their own toolset to provide to mcp_run_python, and make them responsible for not including mcp_run_python itself? Maybe there's another point where we could detect this and raise an error, instead of trying to filter proactively?

DouweM · 2025-07-29T17:31:59Z

pydantic_ai_slim/pydantic_ai/agent.py

        def _set_sampling_model(toolset: AbstractToolset[AgentDepsT]) -> None:
            if isinstance(toolset, MCPServer):
                toolset.sampling_model = sampling_model

        self._get_toolset().apply(_set_sampling_model)

+    def set_mcp_elicitation_toolset(self, toolset_for_elicitation: AbstractToolset[Any] | None = None) -> None:


On second thought, I think we should drop this method as it's too mcp-run-python specific to live on agent. I think it's fine to require the user to pass their own toolset to create_tool_elicitation_callback (and create_auto_tool_injection_callback)

DouweM · 2025-07-29T17:33:43Z

pydantic_ai_slim/pydantic_ai/mcp.py

+    """
+    filtered_toolset = _FilteredToolset(toolset)
+
+    async def elicitation_callback(context: Any, params: Any) -> mcp_types.ElicitResult | mcp_types.ErrorData:


Instead of Any, can we use the actual types from the mcp package's ElicitationFnT definition?

DouweM · 2025-07-29T17:34:45Z

pydantic_ai_slim/pydantic_ai/mcp.py

+        raise ValueError(f'Tool {tool_name} not found in toolset')
+
+    tool = available_tools[tool_name]
+    return await toolset.call_tool(tool_name, tool_arguments, ctx, tool)


Note that this will not perform argument validation. I think we'll need to build a ToolManager and use handle_call

DouweM · 2025-07-29T17:38:58Z

pydantic_ai_slim/pydantic_ai/mcp.py

+    async def elicitation_callback(context: Any, params: Any) -> mcp_types.ElicitResult | mcp_types.ErrorData:
+        """Handle elicitation requests by routing to the toolset."""
+        try:
+            tool_execution_data = json.loads(params.message)


Can we use TypeAdapter(ToolCallPart) to validate this, which we can then directly pass through to tool_manager.handle_call (see my other suggestion about using that instead of call_tool)?

yamanahlawat mentioned this pull request Jul 19, 2025

Tool Injection for mcp-run-python: Enable Python code to call back to agent tools with request/response flow #2037

Open

yamanahlawat marked this pull request as draft July 19, 2025 17:27

yamanahlawat force-pushed the features/main branch from 322696f to 393f025 Compare July 19, 2025 18:14

yamanahlawat marked this pull request as ready for review July 20, 2025 10:11

DouweM self-assigned this Jul 22, 2025

DouweM added the awaiting author revision label Jul 22, 2025

yamanahlawat force-pushed the features/main branch from cee4c16 to 7730660 Compare July 25, 2025 13:24

yamanahlawat added 6 commits July 26, 2025 00:10

feat: Added MCP elicitation support for tool injection in mcp-run-python

48675e2

refactor: mcp-run-python main deno

570a379

feat: replace global _mcp_servers with toolsets

614b477

refactor: simplify MCP elicitation by removing tool name mapping comp…

797cd92

…lexity

chore: update MCP package to v1.12.2 and fix merge conflicts

7433802

refactor: optimized mcp and tool elicitation

04124c2

yamanahlawat force-pushed the features/main branch from cf1f4f4 to 04124c2 Compare July 25, 2025 18:42

fix: merge conflict missing elicitation in client

2129a24

AntonGomes mentioned this pull request Jul 28, 2025

Support MCP elicitations both in client and server #2330

Open

DouweM requested changes Jul 29, 2025

View reviewed changes

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

Added MCP elicitation support for tool injection in mcp-run-python #2258

Added MCP elicitation support for tool injection in mcp-run-python #2258

yamanahlawat commented Jul 19, 2025 •

edited by DouweM

Loading

Uh oh!

yamanahlawat commented Jul 20, 2025

Uh oh!

DouweM commented Jul 22, 2025

Uh oh!

yamanahlawat commented Jul 22, 2025

Uh oh!

yamanahlawat commented Jul 25, 2025

Uh oh!

DouweM left a comment

Uh oh!

DouweM Jul 29, 2025

Uh oh!

DouweM Jul 29, 2025

Uh oh!

DouweM Jul 29, 2025

Uh oh!

DouweM Jul 29, 2025

Uh oh!

DouweM Jul 29, 2025

Uh oh!

DouweM Jul 29, 2025

Uh oh!

DouweM Jul 29, 2025

Uh oh!

DouweM Jul 29, 2025

Uh oh!

DouweM Jul 29, 2025

Uh oh!

Uh oh!

		)


		def create_tool_elicitation_callback(toolset: AbstractToolset[Any]) -> ElicitationFnT:

Added MCP elicitation support for tool injection in mcp-run-python #2258

Are you sure you want to change the base?

Added MCP elicitation support for tool injection in mcp-run-python #2258

Conversation

yamanahlawat commented Jul 19, 2025 • edited by DouweM Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

yamanahlawat commented Jul 20, 2025

Uh oh!

DouweM commented Jul 22, 2025

Uh oh!

yamanahlawat commented Jul 22, 2025

Uh oh!

yamanahlawat commented Jul 25, 2025

Uh oh!

DouweM left a comment

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!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

yamanahlawat commented Jul 19, 2025 •

edited by DouweM

Loading