nREPL connection should not be necessary to use ClojureMCP

[bhauman]

The nREPL connection is primarily needed for the clojure_eval tool, while this does provide a lot of value it shouldn't be a hard requriement.

[ericdallo]

@bhauman how hard it is to solve this?
My point: clojureMCP should not fail to start on initialize message if nrepl is not connected or running, this should be something that the clojure_eval tool should respond with a output to LLM saying that.

[bhauman]

The only thing tricky about this is: where do you get the root directory of the project?

ClojureMCP can be launched from anywhere by LLM clients like ClaudeDesktop.

You can't rely on the current directory being the directory of the project. So I've been leaning on the nREPL connection to to discover the working directory of the project. I.e the directory where you launch the nREPL.

None of the tools really work if they don't know where they are in the file system... The tools have built in restrictions that prevent editing files outside of the working directory.

Currently you can provide a :project-dir on the command line and that solves the need for a working directory at the expense of needing to hard code a directory which doesn't work well for Clients like Claude Desktop and perhaps eca.

And there is another thing, we use the nREPL :describe command connection to determine what kind of repl we are connected to so early on when we are creating tools we know if we are connected to Clojure, babaska, scittle, or basilisp. Although this can probably be lazy.

So its very convenient to have the nREPL connection at the start as this information can be passed to tools and they can change their tool descriptions based on the environment etc, we can also choose a different set of tools for a given environment.

That being said if you aren't using the eval tool and you are providing the :project-dir it seems that ClojureMCP shouldn't need an nREPL environment.

A lazy nREPL environment is possible the major need here is to make sure that at the start of the process the working-directory is provided AND the nREPL connection will need a port number to connect to.

[ericdallo]

@bhauman I understand your points, but still, I'd say we should not fail the initialize request, maybe we should check for project root and do the rest of stuff async after initialize?

Also, isn't included in the MCP spec the addRoot support? That's how MCPs know the project root, right? Shouldn't we rely on that?

[bhauman]

The Client doesn't know the root in the case of Claude Desktop, ECA knows the root, Claude Code knows the root, but Claude Desktop does not.

Currently as far as I can tell ClojureMCP needs to at least have the project-directory to be able to send any tool information to the MCP client. It needs read the config.edn file to determine what tools are available.

Yes the MCP protocol allows for dynamic toolListChanged and ToolAdded like events but I've found Clients (Claude Code and Claude Desktop) are not responsive to dynamic changes of the availability of tools. I wish they were but in practice they seem to ignore updates. So if you start with no tools you are lucky if any tool changes show up in the Client later. The config.edn determines the list of tools and I've found for most clients you need to get them that list immediately or you will have problems with tool availability in general.

So for now AFAIK the project-dir is needed at startup which is not unreasonable. Perhaps if the clients support latent tool additions and changes we can start the server. We should be able to latently start the nREPL but we need to know the project-dir... but if the Clients have improved and are now robustly responding to tool list change events then a more sophisticated server that can be configured at runtime can be built.

[bhauman]

Let me reflect on this some more.

[ericdallo]

Got it, not a trivial issue indeed

[licht1stein]

How about a forced decision about the root directory based on the location of the .clojure-mcp directory with config.edn in it?

[ClemRz]

Having the same issue with Amazon Q (plugin in IntelliJ IDEA) (Claude Sonnet 4):
Me:

Inspect current project

Amazon Q:

I'll inspect the current Clojure project to understand its structure, dependencies, and configuration.

Inspecting the Clojure project structure to understand dependencies, source paths, and configuration

clojure_inspect_project

Error
Parameters

{}
Error

clojure_inspect_project tool input validation failed: root: must have required property 'explanation'