The Best
Tool Discovery
for AI Agents

A context window is the total token budget an LLM has for a single request — including system prompt, tool definitions, messages, and the response. Loading 200+ tool schemas can consume 70,000+ tokens before the agent starts reasoning. OpenAI caps requests at 128 tool definitions. Tool discovery solves this by loading only the tools needed per query. See our guide on MCP token optimization.

Semantic tool search uses a fine-tuned bi-encoder model to match natural language queries against tool descriptions by meaning, not keywords. For example, the query "enroll new hire in training" matches workday_create_learning_enrollment even with zero word overlap — picking the correct tool on the first try 92.8% of the time across 14 unseen connectors.

BM25 matches exact terms — it picks the correct tool on the first try only 14% of the time in our BM25 vs TF-IDF hybrid search evaluation. Semantic search understands intent: different words, same meaning. StackOne's model finds the right tool in its top 5 results 84% of the time, versus 21% for BM25, across 9,340 actions. The SDK's auto mode uses semantic first and falls back to BM25+TF-IDF locally if the API is unreachable.

Add tool-mode=search_execute to your StackOne MCP URL. Your agent will see two tools — tool_search and tool_execute — instead of hundreds of individual definitions. Remove the parameter to revert. No migration, no code changes.

Yes. Add tool-mode=search_execute to your StackOne MCP URL and paste it into any MCP client — Claude, ChatGPT, Cursor, Windsurf, VS Code, Zed. The agent sees tool_search and tool_execute instead of hundreds of definitions. No SDK or code changes needed on the client side.

If your agent uses fewer than 10 tools, static definitions work fine — context cost is negligible and tool availability is deterministic. Tool search adds value when connecting multiple systems where tool count exceeds what fits in the context window, or when you need semantic search across connectors.

The semantic API requires network connectivity — the local BM25+TF-IDF fallback is less accurate at roughly 65% vs 90%. Multi-account routing for the same connector type picks the first match. The tool-mode parameter is set at URL level, not per-request. Code-mode execution (sandbox) is a separate future capability.