Kit by Speakeasy launches open-source coding runtime

Kit’s strongest idea is treating coding agents as composable infrastructure rather than editor-specific assistants, though its performance claims need independent validation.

• –The `compose` model batches shell commands, edits, tests, retries, and parallel work into fewer model round trips.
• –ACP separates the runtime from the client, letting Kit work across compatible editors and orchestrate Claude, Codex, Cursor, or other harnesses.
• –A static MIT-licensed binary with MCP, A2A, Agent Skills, and plugin support makes Kit unusually portable for local or self-hosted workflows.
• –Kit explicitly provides no sandbox or permissions framework, so it is a runtime—not a security boundary—and must run inside a trusted environment.
• –Structured subagent outputs are supported, but schema mismatches return raw text and advance the session, leaving validation policy to developers.