100% failures during tool use

[dontevenblink]

I have a green light in my cursor MCP settings, but I get 100% failure on tool use. All tool calls result in an error like this in the MCP settings UI:

In Claude 3.7, I get a mixture of that error or just no successful tool call at all. With GPT-4o I get tool calls which always produce the above error, but sometimes return the results of tool use anyway. Every error I get requires a reset in the Cursor MCP UI, even when I get some of that seemingly successful return with GPT-4o.

PS -- thanks for working on this, it seems very cool!

[ProgramComputer]

I changed the logging. Can you try again?

[dontevenblink]

That change has fixed having to disable/enable the MCP in the Cursor settings after every error, but I am still having near zero success. In the cursor Agent I'm repeatedly getting errors like this with Claude 3.7:

For many tools, Claude Sonnet will say it will attempted but doesn't indicate that it actually attempted a tool call:

or errors like this in GPT-4o:

The only successful tool call I've made so far has been APOD in GPT-4o.

[ProgramComputer]

Nasa Space Sounds has an access error and I had to remove it. Mars Rover Photos should be working. Let me know of any further issues or additional APIs that are missing.

What's happening in APOD is that the client was calling with no parameters and it should be fine now.

Feel free to suggest prompts that may be useful

[dontevenblink]

Nasa Space Sounds has an access error and I had to remove it. Mars Rover Photos should be working. Let me know of any further issues or additional APIs that are missing.

Feel free to suggest prompts that may be useful

Sorry, perhaps I have been unclear: I'm getting 100% failure, zero success with any of the tools, in Claude 3.7 in Cursor. Sometimes you can see the tool gets a response, but Claude is unaware of it:

This is followed by nothing and no awareness of that response.

In GPT-4o I get some success, but many failures. See below:

[ProgramComputer]

Are you passing in the NASA_API_KEY?

If you can give specific queries, I can test.

If time is the issue, you can look at https://github.com/modelcontextprotocol/servers/tree/main/src/time.

[dontevenblink]

I am passing my API key like this, yes:

I've also got some results, so I know the key is working. I really don't understand how you're not having these problems yourself, though. Here is an example prompt with Cursor 0.47.8, and Claude 3.7:

Notice no tool call is even attempted. If i try to enforce sending a date in the prompt, I will get something:

edit: I'm getting some better results by demanding the arguments the tools need in my prompts. I would have thought more information on how the model needs to use the tool would be baked in.

[ProgramComputer]

I tried the Claude Desktop client and had no issues. Perhaps this is specific to clients

Can you tell me which prompts you used specifically along with package version? And the errors you saw?