implement R as an MCP client #34

simonpcouch · 2025-05-28T20:35:08Z

A proof of concept to close #29. ~~Marking as draft as 1) this is a very scrappy implementation at the moment and 2)~~ we would definitely want to revisit naming before merging. Also worth noting that this only implements the stdio protocol and not (yet) the http one.

With the GitHub MCP server mentioned in the docs configured:

library(acquaint)
  
ch <- ellmer::chat_anthropic()
#> Using model = "claude-3-7-sonnet-latest".

# currently have the github MCP server configured
ch$set_tools(mcp_tools())

ch$chat("What issues are open on posit-dev/acquaint?")
#> I'll help you find the open issues on the posit-dev/acquaint repository. Let me
#> search for that information.
#> ◯ [tool call] list_issues(owner = "posit-dev", repo = "acquaint", page = 1L,
#> perPage = 100L, state = "open", direction = "desc", since = "", sort =
#> "created")
#> ● #> {"jsonrpc":"2.0","id":3,"result":{"content":[{"type":"text","text":"[{\"i…
#> Here are the currently open issues on the posit-dev/acquaint repository:
#> 
#> 1. **Issue #29: Supporting "the other direction"**
#>    - Opened by: simonpcouch
#>    - Created on: May 20, 2025
#>    - This issue discusses adding functionality to allow ellmer chats to act as 
#> a client to other existing MCP servers, which would be the opposite direction 
#> of the current functionality (where acquaint allows clients to interface with R
#> sessions via an MCP server).
#> 
#> 2. **Issue #8: Does the server work with vscode mcp client?**
#>    - Opened by: sounkou-bioinfo
#>    - Created on: April 28, 2025
#>    - This issue reports problems using the acquaint server with VSCode Copilot.
#> The user is experiencing issues where VSCode gets stuck when trying to use 
#> GPT-4.1 or Claude 3.7 with the acquaint MCP server on Ubuntu 24.04.
#>    - This issue has 3 comments.
#> 
#> There are 2 open issues in total on the posit-dev/acquaint repository.

^{Created on 2025-05-28 with reprex v2.1.1}

The first portion of this test should run with the configured GitHub secrets, and we can enable the second portion with a `live-api.yaml`

In attempt to resolve: ``` ── Error ('test-client.R:25:3'): mcp_tools works ─────────────────────────────── <c_error/rlib_error_3_0/rlib_error/error/condition> Error in `c("process_initialize(self, private, command, args, stdin, stdout, ", " stderr, pty, pty_options, connections, poll_connection, env, " )`: ! Native call to `processx_exec` failed Caused by error in `chain_call(c_processx_exec, command, c(command, args), pty, pty_options, ...`: ! cannot start processx process '' (system error 2, No such file or directory) @unix/processx.c:613 (processx_exec) ```

* Use `parse_json()` over `fromJSON()` to avoid conversion to data frame * Handle every ellmer object type, including arrays/objects, and actually note if they're required or not Still need to find a way to not set defaults in the function definition if the argument is required.

This Action fails on macOS and the post action runs indefinitely when that happens.

simonpcouch · 2025-06-06T17:49:09Z

Naming things feels really tricky here. Here's one proposal, but I'm not that excited about it:

R as...	Current	New
Server	`mcp_server()`	`mcp_server()`
	`mcp_session()`	`mcp_session()`
	`mcp_set_tools()`	`mcp_server_tools()`
Client	`mcp_tools()`	`mcp_client_tools()`
	`set_mcp_config()`	`mcp_client_config()`

In the "New" case mcp_*_tools() is both a getter and a setter :/
- This might be made clearer by exporting a getter for the server tools?
I'd anticipate that what is currently mcp_tools() will get a lot of traffic, whereas mcp_set_tools() will get comparatively little.

simonpcouch

Some specific bits where your intuition is appreciated, @gadenbuie :)

simonpcouch · 2025-06-06T18:59:47Z

R/client.R

+#' @returns
+#' * `mcp_tools()` returns a list of ellmer tools that can be passed directly
+#' to the `$set_tools()` method of an [ellmer::Chat] object.
+#' * `set_mcp_config()` returns the path to the new MCP config file, invisibly.


Do you feel these two entry points are a reasonable interface to this functionality? i.e. 1) configure connections in the same way as Claude Desktop via a .json file, and then 2) collect them all as a $set_tools()-compatible list?

simonpcouch · 2025-06-06T19:01:42Z

R/client.R

+server_as_ellmer_tools <- function(server) {
+  tools <- server$tools$tools
+
+  tools_out <- list()


This might be a browser()-worthy point, to poke at the tools object. Is this a reasonable way to go from that json to a list of ellmer tools? Does this duplicate anything in ellmer / fail to make use of relevant functions from that package?

You could potentially build on the contents_record() and contents_replay() helper methods proposed by Barret over in tidyverse/ellmer#503. In that world, I think you'd want to create an McpToolDef class with ellmer:::ToolDef as its parent and then you'd have custom methods for each. (Or anyway, you'll probably want to keep contents_record() and contents_replay() in mind because their intended to support serializing a Chat object for bookmarking in shinychat.)

simonpcouch · 2025-06-06T19:04:37Z

NAMESPACE

@@ -3,4 +3,6 @@
 export(mcp_server)
 export(mcp_session)
 export(mcp_set_tools)
+export(mcp_tools)
+export(set_mcp_config)


Thoughts on better names certainly appreciated, though maybe this has much to do with #35.

I think mcp_tools() needs a verb. Is it a counterpart to mcp_set_tools()? If so, then maybe mcp_get_tools()?

I think you're right in #34 (comment) and I'd advocate for

Sticking with mcp_server_ and mcp_client_ as prefixes

Calling mcp_session() something else like mcp_server_add_session() or mcp_server_connect_session() or just mcp_server_connect() (I think you could get away with this last one if there's a hard line between mcp_server_ and mcp_client_.

Separating mcp_{serer,client}_tools() into _get_tools() and _set_tools(), but only if both actions are relevant.

btw, I think you should also differentiate between mcp_server_ as in R-as-a-server and other external or remote MCP servers, which might use remote as the name. I think this would mean the$mcp_servers would be the$mcp_remote_servers or just the$mcp_remotes. Just in general, anything you can do to remove the cognitive overhead of "server" appearing all over the place would be great

R/client.R

`mcp_client_config()` is a getter, and its output can be adjusted by configuring an option. Users can just supply a different path to `mcp_tools()` to fetch a different set of tools. Uses the newer suggested name since `set_*()` no longer makes sense, but otherwise doesn't make changes to naming.

simonpcouch added 3 commits May 28, 2025 15:29

implement R as an MCP client

7c25481

devtools::document()

1935361

address dependency check WARNINGs

ddb5fa7

simonpcouch mentioned this pull request May 29, 2025

New Feature Request : Integration of ellmer with MCP (Model Context Protocol) Server tidyverse/ellmer#375

Open

simonpcouch added 3 commits May 29, 2025 08:51

revamp docs

47ae699

more resilient config parsing

78533e7

test with GitHub MCP + 4.1-mini

d77eb5c

The first portion of this test should run with the configured GitHub secrets, and we can enable the second portion with a `live-api.yaml`

simonpcouch mentioned this pull request Jun 2, 2025

distinguishing acquaint and btw #35

Open

simonpcouch added 7 commits June 6, 2025 08:07

merge upstream

e7f7544

suggest withr

0707d32

refine skips, resolve error when no servers configured

f43fa79

simplify + remove carrier dependency

ce7d760

skip docker install on macOS runners

d1fcae4

This Action fails on macOS and the post action runs indefinitely when that happens.

revisit docs [no ci]

0700242

simonpcouch requested a review from gadenbuie June 6, 2025 18:30

simonpcouch marked this pull request as ready for review June 6, 2025 18:30

simonpcouch commented Jun 6, 2025

View reviewed changes

gadenbuie reviewed Jun 6, 2025

View reviewed changes

R/client.R Outdated Show resolved Hide resolved

simonpcouch added 2 commits June 11, 2025 13:17

correct error message + test error header

c237221

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

implement R as an MCP client #34

implement R as an MCP client #34

Uh oh!

simonpcouch commented May 28, 2025 •

edited

Loading

Uh oh!

simonpcouch commented Jun 6, 2025

Uh oh!

simonpcouch left a comment

Uh oh!

simonpcouch Jun 6, 2025

Uh oh!

simonpcouch Jun 6, 2025

Uh oh!

gadenbuie Jun 6, 2025

Uh oh!

simonpcouch Jun 6, 2025

Uh oh!

gadenbuie Jun 6, 2025

Uh oh!

gadenbuie Jun 6, 2025

Uh oh!

gadenbuie Jun 6, 2025

Uh oh!

Uh oh!

Uh oh!

implement R as an MCP client #34

Are you sure you want to change the base?

implement R as an MCP client #34

Uh oh!

Conversation

simonpcouch commented May 28, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

simonpcouch commented Jun 6, 2025

Uh oh!

simonpcouch left a comment

Choose a reason for hiding this comment

Uh oh!

simonpcouch Jun 6, 2025

Choose a reason for hiding this comment

Uh oh!

simonpcouch Jun 6, 2025

Choose a reason for hiding this comment

Uh oh!

gadenbuie Jun 6, 2025

Choose a reason for hiding this comment

Uh oh!

simonpcouch Jun 6, 2025

Choose a reason for hiding this comment

Uh oh!

gadenbuie Jun 6, 2025

Choose a reason for hiding this comment

Uh oh!

gadenbuie Jun 6, 2025

Choose a reason for hiding this comment

Uh oh!

gadenbuie Jun 6, 2025

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

simonpcouch commented May 28, 2025 •

edited

Loading