Most MCP servers just wrap CRUD JSON APIs into tools — I did it too with scim-mcp and garmin-mcp-app. It works, until you realize a tool call dumps 50KB+ into context.
MCP isn't dead — but we need to design MCP tools with the context window in mind.
That's what code mode does. The LLM writes a small script, the server runs it in a sandbox against the raw data, and only the script's compact output enters context.
Inspired by Cloudflare's Code Mode, but using a local sandboxed runtime instead of a remote one — no external dependencies, isolated from filesystem and network by default.
Works best with well-known APIs (SCIM, Kubernetes, GitHub, Stripe, Slack, AWS) because LLMs already know the schemas — they write the extraction script in one shot.
Copy-paste this into any AI agent inside your MCP server project:
Add a "code mode" tool to this MCP server. Code mode lets the LLM write a processing
script that runs against large API responses in a sandboxed runtime — only the script's
stdout enters context instead of the full response.
Steps:
1. Read the codebase. Identify which tools return large responses.
2. Pick a sandbox isolated from filesystem and network by default:
- TypeScript/JS: `quickjs-emscripten`
- Python: `RestrictedPython`
- Go: `goja`
- Rust: `boa_engine`
3. Create an executor that injects `DATA` (raw response as string) into the sandbox,
runs the script, captures stdout.
4. Create a code mode MCP tool accepting `command`, `code`, and optional `language`.
5. Create a benchmark comparing before/after sizes across realistic scenarios.
Walk me through your plan before implementing. Confirm each step.
Also inspired by claude-context-mode, but instead of running code in a child process, this approach uses a fully isolated sandbox (WASM or embedded interpreter) with no filesystem or network access by default.