Does Marvin support MCP?

[FreeDemon]

Does Marvin support MCP?

[zzstoatzz]

hi @FreeDemon - because we're downstream of pydantic-ai, technically yes

is there some marvin-specific MCP functionality you're looking for?

[slavaGanzin]

Have the same question.

So for example I want to use https://github.com/PrefectHQ/marvin/blob/da23cb8485d1e9b1668801cdfeef886333a8d32a/examples/you_have_been_goosed/goose_em.py in my Marvin-controlled chat bot. How this could be done? How I add it as a tool, or what?

[FreeDemon]

hi @FreeDemon - because we're downstream of pydantic-ai, technically yes

is there some marvin-specific MCP functionality you're looking for?

yes, I have a MCP server URL based on SSE provided by Zapier, and I would like to know how it can be used in Marvin?

[zzstoatzz]

i would recommend using pydantic_ai (included when installing marvin) support for MCP at this time - it's not immediately clear to me what marvin specifically would offer.

I often use pydantic-ai directly when I don't need the additional abstraction/ease of use that marvin offers

Marvin-controlled chat bot. How this could be done? How I add it as a tool, or what?

@slavaGanzin example you linked is an MCP server that you can add to a pydantic_ai Agent like the example linked above

how it can be used in Marvin?

I would be happy to think more about this if you have more detail on what you're looking for

[AndrewHannigan]

I'm +1 to this request, mainly because I think the marvin abstractions are really nice and I'd like to build apps that use MCP with it. One example I currently have is that I want to write a simple Agent with marvin that performs file editing. I'm interested in this MCP server specifically. But currently to do so I'd have to breakout into pydantic-ai as you suggested @zzstoatzz. Which I can do easily enough, but then I don't get to use marvin and all the nice abstractions that come with it 🤖

A lightweight wrapper around MCPServerStdio could do the trick. Something like:

def create_mcp_stdio_tools(command: str, args: list[str]) -> list[Callable]:
    # Initialize MCPServerStdio 
    
    # Create a callable for each tool in the MCP server

It would act like a closure, and the server would continue running within the scope of the create_mcp_stdio_tool call after the list of Callables is returned. I think marvin inspects the docstring of the Callable, so this could be taken from the tool's description in the MCP server.

[AndrewHannigan]

I'd be happy to take a stab at a PR. Interested in feedback first though on whether that approach makes sense.

[zzstoatzz]

great @AndrewHannigan I'd be happy to review.

do we even need a wrapper or do we just need to update agents.py to allow passing an mcp_servers kwarg to the underlying pydantic_ai.Agent? i haven't had time to look closely at this yet, so just asking

[AndrewHannigan]

@zzstoatzz I haven't done a deep dive on marvin codebase since the replatform on Pydantic AI so I'll take a look. Any callbacks or other special logic that marvin attaches to tool calls wouldn't be applied to mcp calls, is that okay in your view?

[AndrewHannigan]

Here's one such callback, on_tool_call. There's nice features we get with marvin tool use like the PrintHandler's terminal formatting. So do I wonder if it's worth wrapping. It would also keep pydantic ai abstracted from user which is nice for future proofing.

If we pass directly to pydantic AI's mcp_servers, how would PrintHandler show these calls? A newon_mcp_server_tool_calls method in Handler?

[zzstoatzz]

ahh yes, true - that sounds reasonable to me!

[zzstoatzz]

still some things to clean up but #1129

» uv run examples/agent_mcp.py
[05/01/25 11:08:04] INFO     marvin._internal.integrations.mcp: Discovered a total of 1 MCP tools.                                                                                mcp.py:181
╭─ Agent "MCP Calculator" (5c49249d) ──────────────────────────────────────────────────────────────╮
│                                                                                                  │
│  Tool:    run_python_code                                                                        │
│  Status:  ✅                                                                                     │
│  Input    {                                                                                      │
│               'python_code': 'from datetime import datetime\n\nstart_date = datetime(2000, 1,    │
│           1)\nend_date = datetime(2025, 3, 18)\n\ndays_difference = (end_date -                  │
│           start_date).days\ndays_difference'                                                     │
│           }                                                                                      │
│  Output:  [                                                                                      │
│               TextContent(                                                                       │
│                   type='text',                                                                   │
│                   text='<status>success</status>\n<return_value>\n9208\n</return_value>',        │
│                   annotations=None                                                               │
│               )                                                                                  │
│           ]                                                                                      │
│                                                                                                  │
╰──────────────────────────────────────────────────────────────────────────────────── 11:08:05 AM ─╯

╭─ Agent "MCP Calculator" (5c49249d) ──────────────────────────────────────────────────────────────╮
│                                                                                                  │
│  Tool:    Mark Task 6ccfa405 ("How many days are between 2000-01-01 and...") successful          │
│  Status:  ✅                                                                                     │
│  Result   9208                                                                                   │
│                                                                                                  │
╰──────────────────────────────────────────────────────────────────────────────────── 11:08:08 AM ─╯
9208

[zzstoatzz]

hello @FreeDemon @slavaGanzin @AndrewHannigan - thanks all for your input thus far. i have an alpha release out (3.1.0a0) with MCP support like this. If you get the chance to try it out, feel free to let me know what should be improved / changed or if you see unexpected behavior (or open a PR directly!)

did this as an alpha release for now to gather feedback / take time to test it more practically. and maybe we ought to expose the mcp_servers kwarg in more places to allow MCP tool use for the high level utils

here's the PR that adds the initial support

[AndrewHannigan]

Awesome, thanks for adding this so quickly! Fyi I'm getting this error but it seems like it is a pydantic issue not marvin.

andrewhannigan:project/ (master) $ uv run test_marvin_mcp.py                                                                                                                                                                                                                        [13:42:44]
Reading inline script metadata from `test_marvin_mcp.py`
--- Running task: ---
Get the latest commit hash from this repo and report how many characters long it is. Then, report the square root of that number and write a 'summary.txt' based on your work.
--------------------

--- Starting Agent Run ---
error: JSR package manifest for '@pydantic/mcp-run-python' failed to load. Import 'https://jsr.io/@pydantic/mcp-run-python/meta.json' failed.

The script I'm running is the same as yours except I added this at top and ran with uv run.

# /// script
# dependencies = [
#     "marvin==3.1.0a0",
#     "pydantic-ai",
# ]
# ///

[AndrewHannigan]

Still having issues, I reproduced with a simplified example

Script:

# /// script
# dependencies = [
#     "marvin==3.1.0a0",
#     "pydantic-ai",
# ]
# ///

import asyncio

from pydantic_ai.mcp import MCPServerStdio
from rich import print as pprint

from marvin.agents import Agent

# Requires Deno: `deno install -A ... jsr:@pydantic/mcp-run-python`
run_python_server = MCPServerStdio(
    command="deno",
    args=["run", "-A", "jsr:@pydantic/mcp-run-python", "stdio"],
)

linus = Agent(
    name="Tester",
    instructions="Use the available tools as needed to accomplish the user's goal.",
    mcp_servers=[run_python_server],
)


async def main():
    task = 'Use python to return the output of `hash("hello world")`'
    pprint(f"--- Running task: ---\n{task}\n" + "-" * 20)
    pprint("\n--- Starting Agent Run ---")
    result = await linus.run_async(task)
    pprint("\n--- Final Result ---")
    pprint(result)


if __name__ == "__main__":
    asyncio.run(main())

Result

test HEAD % uv run ~/Downloads/test_marvin_mcp.py
--- Running task: ---
Use python to return the output of `hash("hello world")`
--------------------

--- Starting Agent Run ---
[05/06/25 18:36:24] ERROR    marvin._internal.integrations.mcp: Error calling MCP tool 'run_python_code': 'str' object has no attribute 'content'                                                      mcp.py:100
                             ╭────────────────────────────────────────────────────────────────── Traceback (most recent call last) ──────────────────────────────────────────────────────────────────╮
                             │ /Users/andrewhannigan/.cache/uv/archive-v0/XxToTiPMqKW_Udwr2U44l/lib/python3.11/site-packages/marvin/_internal/integrations/mcp.py:85 in _mcp_tool_wrapper            │
                             │                                                                                                                                                                       │
                             │    82 │   │   result: CallToolResult = await _mcp_server.call_tool(                                                                                                   │
                             │    83 │   │   │   tool_name=tool_name, arguments=kwargs                                                                                                               │
                             │    84 │   │   )                                                                                                                                                       │
                             │ ❱  85 │   │   result_content = result.content                                                                                                                         │
                             │    86 │   │                                                                                                                                                           │
                             │    87 │   │   await _orchestrator.handle_event(                                                                                                                       │
                             │    88 │   │   │   ToolResultEvent(                                                                                                                                    │
                             ╰───────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────╯
                             AttributeError: 'str' object has no attribute 'content'
╭─ Agent "Tester" (47ef7f0b) ──────────────────────────────────────────────────────────────────────╮
│                                                                                                  │
│  Tool:    run_python_code                                                                        │
│  Status:  ✅                                                                                     │
│  Input    {'python_code': "result = hash('hello world')\nresult"}                                │
│  Output:  "Error calling tool run_python_code: 'str' object has no attribute 'content'"          │
│                                                                                                  │
╰────────────────────────────────────────────────────────────────────────────────────  6:36:21 PM ─╯
╭─ Agent "Tester" (47ef7f0b) ──────────────────────────────────────────────────────────────────────╮
│                                                                                                  │
│  Tool:    run_python_code                                                                        │
│  Status:  ✅                                                                                     │
│  Input    {'python_code': "result = hash('hello world')\nresult"}                                │
│  Output:  "Error calling tool run_python_code: 'str' object has no attribute 'content'"          │
│                                                                                                  │
╰────────────────────────────────────────────────────────────────────────────────────  6:36:21 PM ─╯
╭
...

[zzstoatzz]

weird, running the same example on main seems to work fine

uv run --isolated --with pydantic-ai examples/agent_mcp.py

but running yours with

pbpaste | uv run -

does reproduce 🤨

will take a look! thanks for trying it out

[zzstoatzz]

aha I think its a new version thing

» uv run --isolated -U --with pydantic-ai examples/agent_mcp.py

this repros

[zzstoatzz]

ok #1133 with your example @AndrewHannigan

» pbpaste | MARVIN_LOG_LEVEL=INFO uv run --with git+https://github.com/PrefectHQ/marvin.git@refs/pull/1133/head -
    Updated https://github.com/PrefectHQ/marvin.git (7cf4b389cc53e8f1465f598d3f52e6b2963d7caf)
      Built marvin @ git+https://github.com/PrefectHQ/marvin.git@7cf4b389cc53e8f1465f598d3f52e6b2963d7caf
Installed 84 packages in 500ms
--- Running task: ---
Use python to return the output of `hash("hello world")`
--------------------

--- Starting Agent Run ---
╭─ Agent "Tester" (6d223d90) ──────────────────────────────────────────────────────────────────────╮
│                                                                                                  │
│  Tool:    run_python_code                                                                        │
│  Status:  ✅                                                                                     │
│  Input    {'python_code': 'hash("hello world")'}                                                 │
│  Output:  '<status>success</status>\n<return_value>\n777879991\n</return_value>'                 │
│                                                                                                  │
╰────────────────────────────────────────────────────────────────────────────────────  8:04:49 PM ─╯
╭─ Agent "Tester" (6d223d90) ──────────────────────────────────────────────────────────────────────╮
│                                                                                                  │
│  Tool:    Mark Task 8dfeb555 ("Use python to return the output of `hash...") successful          │
│  Status:  ✅                                                                                     │
│  Result   777879991                                                                              │
│                                                                                                  │
╰────────────────────────────────────────────────────────────────────────────────────  8:04:52 PM ─╯

--- Final Result ---
777879991

which is now released as 3.1.0a1

[AndrewHannigan]

This is working for me now too 🙌

The release title 😆 Tim Robinson would be proud