Arcade launches MCP tool discovery for enterprise AI agents

Arcade.dev launched Tool Recommendation on October 15, giving AI agents a way to retrieve a short list of relevant tools instead of loading every available tool definition into their context. The company said the feature is available to all Arcade Cloud users, with a separate rollout underway for enterprise customers.

Tool Recommendation operates through Model Context Protocol gateways. Once it is enabled, an MCP client receives a small, fixed set of Arcade functions instead of the gateway’s full allowed-tool list. The agent describes its task to Arcade.SelectTools. Arcade then uses embedding and ranking models to match the request against tool names and descriptions, returning ranked candidates and their input schemas. The agent chooses from the shortlist and calls Arcade.UseTool to execute the selection. Arcade’s documentation says the selection call returns up to five tools per task by default; the REST interface can raise the limit to 20.

That discovery step does not widen permissions. Arcade says it searches only the tools allowed by the gateway, while execution remains subject to user authorization, agent permissions and policy checks in its actions runtime.

Arcade’s headline efficiency figures come from different measurements. Its claim of up to 37% lower token cost is based on a company model that compares Tool Recommendation with an assistant carrying a fixed list of 100 tools. The calculation counts input cost, excludes output tokens and reaches its maximum in single-request sessions. Arcade estimates savings of about 20% for what it defines as a typical conversation, falling to 5% to 7% when large tool results remain in the conversation. The company’s claim of up to 91% fewer tokens covers tool definitions and recommendation responses in a modeled 10-turn session. It excludes the conversation and tool results, so it is not a whole-session reduction.

A separate company-run test produced a smaller figure that included both input and output. Across 408 paired sessions spanning 64 workflows with Codex and Claude Code, Arcade said the total token bill fell 7.6%, with a 95% interval of 2.5% to 12.6%, while task completion was unchanged. The company reported a 9.7% reduction for Claude Code and 5.5% for Codex, although the Codex confidence interval included zero. The tests have not been independently reproduced.

For smaller catalogs, the economics can reverse. Arcade’s model found that a fixed list was cheaper below roughly 25 to 28 tools. Its limitations documentation also says gateways with fewer than about 10 tools gain little because discovery adds another round trip. Recommendation quality depends on tool names and the first line of their descriptions; Arcade warns that vague or nearly identical descriptions can push the correct tool down the ranking.