MCP/Code Mode

• MCP/Code Mode

  • An execution strategy built on top of MCP (not part of the MCP spec itself) that emerged as a popular pattern in late 2025
  • Instead of the model calling individual tools one at a time, the model writes a short program that calls many tools programmatically; only the final result returns to the model context

  • Traditional MCP vs Code Mode

    • Traditional MCP

      • MCP server exposes 50–500 tools
      • The LLM sees all tool schemas upfront (large context cost)
      • The LLM decides which tool to call, calls it, sees the result, then decides the next call
      • Intermediate results flow through the model context at every step

    • Code Mode

      • MCP server exposes a small set of "meta tools"
      • The LLM discovers available capabilities, then writes code (Python, TypeScript, etc.)
      • That code executes in a sandbox and calls the underlying MCP tools programmatically
      • Only the final result is returned to the model — intermediate data stays in the sandbox

    • Example

      • Instead of six round-trips through the model (search_github, get_issue, get_comments, get_pr, get_diff, summarize), the model writes:
        • python

          issues = github.search_issues(...)
          comments = github.get_comments(...)
          prs = github.get_prs(...)

        • and executes it in one step

  • Why it matters

    • Token efficiency — some implementations report 90%+ reductions in context devoted to tool definitions; an entire API can fit in ~1000 tokens
    • Better orchestration — LLMs are generally stronger at writing sequential code than emitting long chains of tool calls, so complex workflows become more reliable
    • Context hygiene — intermediate data stays inside the sandbox instead of polluting the model's working context
    • Multi-server composition — easier to orchestrate across many MCP servers simultaneously

  • Mental model

    • MCP is the protocol. Code Mode is an execution strategy on top of MCP.
    • Analogous to: calling a REST API endpoint-by-endpoint vs. generating a typed SDK and writing code against it
    • MCP servers increasingly look less like collections of tools and more like dynamically generated libraries that agents can program against

  • Implementations and references

    • Cloudflare/MCP — "Code Mode: the better way to use MCP" and "give agents an entire API in 1000 tokens"
    • FastMCP — Code Mode docs
    • Bifrost AI Gateway — Code Mode
    • How Code Mode Builds on MCP for Agent Tooling — Nordic APIs
    • Code Mode with Skills — Navendu Pottekkat

  • See also

    • MCP/Server
    • MCP/Gateway/Enterprise