Best way to do tool/source discovery when adding Executor to an agent configuration?

[wesmith4]

Problem

When I add Executor as a configured MCP server to my agent environment with the proper config, the agent session is not discovering the tools/sources that I have added there.

For example, I've added the OpenAPI spec for the Eppo REST API as a source in my executor config, but when I ask Claude Code about my Eppo experiments, it is not aware that it has tools available for accessing these resources through executor.

Request

Implement & document some proper spec for how AI agents with executor installed and configured to proactively discover the namespaces/tools that are available through Executor.

[RhysSullivan]

Do you tell it to use Executor? Claude Code lazy loads MCP tools resulting in this, I believe if you tell it along the lines of '... use executor' it triggers a regex getting the tool description actually into the model

Looks like it cuts off the tool description a well so yeah more work to do here

[wesmith4]

Yeah, when I've told it to use executor, it does do the discovery, by way of those namespaces listed in the tool description and then executing the typescript. My main question would be what would be the cleanest way to make the AI agent understand what namespaces/tools are being made available via executor as a part of its context/tool search upon launch.

As I think you're getting at in your comment, perhaps some optimization of the positioning of namespace list in the tool description would be a good way to start.

End state ofc, we want this to match performance of the native tool discovery / search if I were to configure each MCP/toolset individually in the .mcp.json or agent skills

[wesmith4]

actually, perhaps a separate tool in the main executor MCP completely focused on returning a list of namespaces would go a long way here to help Claude et. al more effectively find the underlying tools exposed through executor

[RhysSullivan]

Closing: the execute tool's description now includes a live per-session integration inventory, and per-integration search tools plus skills({name:"execute"}) cover discovery. That's the shape this asked for.¹

Footnotes

1. Sent from my Claude ↩