Vercel Limits Exposure to MCP Tools to 15 Schemas

Vercel offers a coding agent that restricts the surface of MCP tools presented to the model. The exposure stabilizes at fifteen function schemas per cycle thanks to a compiled registry and on-demand dereferencing. This design comes with a search limit and aims to avoid inappropriate tool choices identified with traditional approaches.

Bounded Exposure: 15 Schemas and 5 Search Results

Vercel's agent limits the model's exposure to MCP tools to fifteen function schemas per cycle, even if multiple MCP servers are installed. Each capacity request can only return five search results. This approach, compared to the on-demand dereferencing of OpenAI Codex, favors a stable tool list size and aims to prevent situations where the model, faced with too many options, makes poor choices. The systematic submission of the same fifteen function schemas materializes this framing.

Compiled Registry and On-Demand Dereferencing

The Vercel agent, observed through the Zig-based fx agent, relies on a registry of tools established at compilation and on a search and dereferencing mechanism. MCP tools do not appear in the list of tools visible to the model at the beginning of a cycle. They are discovered and inserted only after the model has selected a tool by its name, which conditions their access.

Limitations of Traditional Agents and Vercel's Response

Traditional coding agents insert all MCP tool schemas into each request, complicating prompts and forcing the model to make a choice among numerous options. If multiple MCP servers are integrated, the increase in alternatives can lead to the selection of tools that are not suitable. To address these challenges, Vercel has launched a new agent that modifies the management of MCP tools. The goal is to simplify the coding experience and eliminate the visible list of MCP tools in favor of a selection triggered by the tool's name.