Precomputing Live demo: MCP How to run it

Answers ready before you ask

An AI agent asks the four demos' files, and each answer costs a few hundred tokens

Loading

Loading SQLite, the MCP server and the four files, about 2 MB.

0
tool calls
0
tokens in the answers
0
tokens reading what each file keeps
0
tokens the raw data would take
–
fewer tokens than the raw data
–
checks against the raw events

Four files, four MCP servers

Each demo's own file, as its policy leaves it once the detail it keeps for a short while has faded. Each gets its own server: the Go code of precomputing serve, built for WebAssembly and running in this page. The agent reads each file's description once.

ServerThe fileAs it standsSizeDescription

Make a call yourself

Pick a server and a tool, change the arguments, and press Run (or Ctrl+Enter, ⌘+Enter on a Mac). Under Any message you can send your own JSON-RPC message and headers, mistakes included, and see what the server says.

A recorded session with a real agent

Loading the recording.

How an agent connects

Outside this page the same server runs in precomputing serve, at /mcp beside the HTTP API, or as precomputing mcp FILE for agents that start local tools. A typical client configuration:


    

The server speaks the current protocol revision, 2026-07-28, and answers clients that open with the older handshake (2025-11-25, 2025-06-18 and 2025-03-26). Every tool only reads. A read token cannot post events. The prototype is not public yet.