Speakeasy launched Kit, an MIT Rust coding-agent runtime that gives the model one compose tool and reports half of Claude Code's input tokens
Kit launched on Product Hunt 2026-09-06 (v0.1.108, MIT, Rust, 32 stars at time of check). Instead of a tool menu, it exposes a single `compose` tool where the model writes a Runlet program that can run shell commands, edit files, run tests, delegate to subagents and return structured data inside one model round trip, which is where the context saving comes from. On Speakeasy's own production tasks from July to August 2026 it reports median 49.7k input tokens per hand-written line against Claude Code's 99.6k, 0.13 min active time against 0.31 min, and 60% fewer user messages per session. It ships as one static binary serving a terminal client, an ACP v1/v2 server over stdio, HTTP/SSE and WebSocket, a bidirectional A2A v1 endpoint, and a subagent orchestrator, and it consumes MCP, Agent Skills and Agent Plugin packages. Treat the benchmark as vendor-reported until someone reproduces it.
↳ Follow the thread