Making the full tool set easier to reach over MCP

[Augusto-Zolet]

Heloooooo! Coming back to the idea of making the full tool set easier to reach over MCP, I went deep on it and worked out the following approach:

The problem (recap)

Over MCP we advertise ~12 core tools; the ~138 specialized ones are only reachable through the delegate_to_specialized_* gateway, which launches a separate sub-agent (a second LLM call) to pick and run the specialized tool, so it only works when an LLM backend is configured. I'd like to reach them directly, without that indirection.

Modes, selectable by a config flag

By default we keep today's behavior; the flag is mcp.tool_exposure_mode:

Mode	What it does	Who it's for
delegate (default)	today's behavior, unchanged	anyone with a sub-agent backend running
direct_flat	advertise all specialized tools	clients with native tool-search (Claude; OpenAI gpt-5.4+)
direct_discovery	small list + a find_tools search tool	everyone, engine-agnostic

1. direct_flat, why keep it (and the honest part)

I checked whether native tool-search "performs better" enough to justify this mode. Honestly: not on retrieval quality, Anthropic's Tool Search Tool is BM25/regex-based, and independent reports suggest a good server-side semantic+BM25 search can match or beat it. So I'm not selling direct_flat as the high-performance mode. Its real value is native integration: capable clients already have an in-context, prompt-cache-friendly search built in, but it only kicks in if the server advertises the full catalog. direct_flat is the server half they need, it's a ~1-line change, and it skips the extra find_tools round-trip for those clients. (Foot-gun: for clients without native search it's the bloat case, so it's explicitly a "capable-client" mode.)

2. direct_discovery (find_tools), the client-agnostic SOTA

• It's basically a search tool that runs on the server. Lots of projects already do exactly this, Docker's MCP Gateway, Cloudflare, IBM ContextForge, FastMCP, so it's a proven, low-risk pattern.
• one tool, find_tools(query, domain?): domain → a precise pre-filter via the existing domain registry; query → ranking over tool name/description/args; returns the top ~5–8 full schemas, plus the available domains and per-domain usage hints. The agent then calls the real tool by name.
• the ~12 core (most-used) tools stay in the list as usual; find_tools only shows up when this mode is on.

On the ranker: I'd start with BM25 (keyword matching), it's lightweight and easy to ship. Semantic search with embeddings would be better, as a natural upgrade later.

The delegate's hints

The delegate also adds small per-domain hints (e.g. for charts in Calc: "you must pass a data_range"). We fold those into the direct modes, into the tool descriptions / the find_tools output, so the agent still gets them.

Thoughts?

[KeithCu]

Hello, thanks for working on this! I don't like the direct_flat idea but whatever, we can have it as a flag for models that can handle it. The find_tools sounds good. It's great to learn from other systems who have had to solve similar problems. Please send the PR and I'll review and check it in. I think keyword search is enough, especially if we tell them the domains to search for.

[KeithCu]

@Augusto-Zolet Besides this, are there any bugs in the latest code you think should be fixed before the next release? This extension is pretty big now (~80kloc not including tests) so I want to make sure it's stable. I added something to report UNO thread violations so if you find any, please let me know or send a PR. I just found a few and fixed them so try on the latest code in Git.

[Augusto-Zolet]

Ok! I'll test it this evening or tomorrow. Next week I'll probably be getting WriterAgent into the hands of a few users at work for real use.

We could also add a bug-report feature, what do you think? We could have a tool the agent can call to report detailed feedback on whatever it ran into, since our users won't have the technical knowledge to describe what's happening themselves. The same tool could cover not just bugs but bad user experience as well.

[KeithCu]

Great idea! Thank you! I added a menu item that pre-fills with basic info. And I also wired up a few message boxes to make it easy to report bugs then and it should copy over the error text although I don't see any errors right now so I'm not sure it works yet. If you find more places to add this helper to make it easy to give feedback on what are not likely user errors, send a PR. Here's the commit: 36941cc

[Augusto-Zolet]

Just opened the PR: #355

[KeithCu]

Awesome! I accepted it and then made a few simplifications in the next commit. We don't need to do fancy search if we just expose the full domain list and tell them to use it. Let me know if you see any issues.