mcp-code-mode

MCP Code Mode

What Code Mode Is

Code Mode is a server-side MCP architecture where the LLM writes code to orchestrate API calls instead of calling individual tools one at a time.

Standard MCP has two scaling problems:

1. Context bloat: Every tool definition loads into the LLM's context upfront. Hundreds of tools = tens of thousands of tokens spent before the first user token. Cloudflare's full API would require 1.17 million tokens as individual tools — exceeding most model context windows entirely.

2. Round-trip overhead: Every tool call is a full inference round-trip. A 20-step task means 20 back-and-forth cycles, each burning latency and tokens on intermediate results that only exist to feed the next step.

Code Mode solves both by collapsing the entire API surface into two meta-tools:

• search (discovery tool): The LLM writes an executable code snippet to find relevant operations on demand, without the full spec ever entering context.
• execute (code execution tool): The LLM writes a script that chains API calls, handles conditionals and pagination, and returns only the final result. The script runs in a secure sandbox server-side.

The result: Typically ~1,000 tokens for discovery setup before task payloads, rather than loading every tool definition up front. One execution call can replace 20+ round-trips. API keys stay server-side and never appear in tool parameters.

Why LLMs write code better than they call tools

LLMs are trained on millions of lines of real-world code. Tool-calling schemas are mostly synthetic training examples. Code is the LLM's native orchestration language — it can express conditionals, loops, error handling, and data transformation that tool-calling schemas cannot.

---

The Two Core Tools

Every Code Mode implementation, regardless of language or framework, converges on the same two-tool surface:

Tool 1: Discovery / Search

Purpose: Let the LLM execute code to find what operations exist without loading the full catalog.