Code Mode is a Cloudflare technique (and @cloudflare/codemode package) that slashes agent token usage by having the LLM write code against a typed SDK — then executing that code safely in an isolated sandbox — instead of calling individual MCP tools one by one. The Cloudflare API MCP server uses Code Mode to expose the entire Cloudflare API through just two tools in under 1,000 tokens, a 99.9% reduction from the 1.17M tokens a traditional per-endpoint MCP server would require.
Why It's in Trial
Code Mode addresses a real and growing problem: as MCP servers expose more tools, the token overhead of listing and invoking them becomes a significant cost and latency multiplier. The technique is novel but already in production on Cloudflare's own MCP server:
- Concrete efficiency gains: 81% token reduction for complex batch operations, 32% for simple tasks (Cloudflare-measured). For large APIs, the gains are effectively unlimited — the Cloudflare API case is 99.9%.
- Open package:
@cloudflare/codemodeis published on npm. The pattern is portable, not Cloudflare-locked. - Secure execution via Dynamic Workers: AI-generated code runs in V8 isolates, not containers — millisecond cold starts, megabyte memory footprint, 100x faster than equivalent container-based sandboxes.
- Presented at MCP Dev Summit North America (April 2-3, 2026): Matt Carey's session "Every API Is a Tool for Agents" demonstrated the technique to the MCP practitioner community at the Linux Foundation event in NYC.
- Broad applicability: The Figma presentation from the summit explicitly frames this as an alternative to raw MCP tool use — useful for any large API surface, not just Cloudflare services.
Stays in Trial (not Adopt) because: the technique is newly published, Dynamic Workers are in open beta, and community adoption outside Cloudflare itself is not yet measured.
When NOT to Use Code Mode
- Small, stable APIs (< ~20 endpoints): the overhead of generating type definitions outweighs the savings
- Agents without a sandboxed execution environment: Code Mode requires a secure runtime to execute LLM-generated code; running arbitrary AI code outside an isolate is unsafe
- Non-TypeScript ecosystems: the package is TypeScript-first; other runtimes need adaptation
- High-trust, low-risk tool use: if your MCP server has 3 tools, raw tool calling is simpler
How It Works
Traditional MCP: [1 tool definition per endpoint] × N endpoints = N × ~100 tokens
Code Mode: TypeScript types (one-time, ~1000 tokens) + execute(code) = fixed cost
The @cloudflare/codemode package:
- Generates TypeScript interfaces from your tool definitions (more token-efficient than OpenAPI for LLMs)
- Exposes two tools to the agent:
search(query)for discovery,execute(code)for action - Executes the returned code in a Dynamic Worker — a V8 isolate instantiated at runtime with only the APIs the code needs
The Dynamic Worker sandbox communicates with host services via Cap'n Web RPC bridges, which operate transparently across the security boundary. Credential injection happens on the way out — agent code never sees secrets directly.
Dynamic Workers Sandbox
| Property | Value |
|---|---|
| Isolation | V8 isolate (same as Cloudflare Workers) |
| Cold start | Milliseconds |
| Memory | Megabytes per isolate |
| vs. containers | ~100x faster startup, ~100x more memory-efficient |
| Security extras | MPK (Memory Protection Keys), Spectre mitigations, dynamic risk-based cordoning |
| Availability | Open beta, paid Workers plan |
| Pricing | $0.002/unique Worker/day (waived in beta) |
MCP vs. Code Mode vs. Skills (Summit Framing)
The Cloudflare MCP Dev Summit session framed three patterns for exposing APIs to agents:
| Pattern | Token cost | Flexibility | When to use |
|---|---|---|---|
| Raw MCP tools | O(N endpoints) | Per-operation | Small, stable APIs |
| Code Mode | Fixed (~1,000 tokens) | Arbitrary composition | Large, dynamic APIs |
| Skills | Near-zero (instructions only) | Opinionated workflows | Teaching agents Cloudflare idioms |
These aren't mutually exclusive — Cloudflare ships all three. The Skills plugin (github.com/cloudflare/skills) complements Code Mode by giving agents intent-level context.
Key Characteristics
| Property | Value |
|---|---|
| Package | @cloudflare/codemode |
| Runtime | Cloudflare Dynamic Workers (V8 isolates) |
| Token reduction | 32–81% typical; up to 99.9% for large APIs |
| Presented | MCP Dev Summit NA, April 2-3, 2026, NYC |
| License | Open-source (Apache 2.0) |
| Requires | Cloudflare Workers paid plan (for Dynamic Workers sandbox) |
Further Reading
- Code Mode: the better way to use MCP (Cloudflare blog)
- Code Mode: give agents an entire API in 1,000 tokens (Cloudflare blog)
- Sandboxing AI agents, 100x faster — Dynamic Workers (Cloudflare blog)
- Cloudflare Code Mode cuts token usage by 81% (WorkOS)
- Codemode API reference (Cloudflare Agents docs)
- MCP Dev Summit North America 2026 (Linux Foundation)