kit - a small coding agent runtime
About Kit
A coding agent runtime from Speakeasy that gives the model a single tool, compose, whose argument is a short program rather than a one-off call. The intent is fewer model round trips: more work completed per request, and less context repeated. It ships a terminal client, an Agent Client Protocol server and an A2A endpoint, speaks ACP v1 and v2 over stdio, HTTP/SSE and WebSocket, and stores sessions as append-only JSONL. Bring your own model, including ChatGPT subscriptions over native OAuth. It is deliberately small and says so: no web UI, no visual debugger, and no permissions framework of its own.
Key Capabilities
One tool, compose, whose argument is a program: fewer round trips per unit of work
Subagents are reusable values you can continue, fork and inspect
Speaks ACP v1 and v2 over stdio, HTTP/SSE and WebSocket, plus A2A and MCP
Sessions are append-only JSONL synced to disk, so long runs survive
Bring your own model, including ChatGPT subscriptions via native OAuth
Small on purpose: the docs state plainly that it is a runtime, not a security boundary
Perfect for
Compare Kit
Community Buzz
Community Buzz
No recent posts yet
We are monitoring X for fresh discussions about Kit.
Similar Tools to Kit
Vibe Kanban
Open-source kanban board for orchestrating AI coding agents. Plan tasks as issues, run Claude Code, Codex, or Gemini CLI in isolated git worktrees, review diffs inline, and open pull requests from one local UI. The company behind it (bloop) shut down in April 2026, but the project lives on as community-maintained open source and local workspaces keep working.
Hermes Agent
Self-hosted autonomous agent from Nous Research with a built-in learning loop. Creates skills from experience, persists memory across sessions, and runs 24/7 on your own hardware with sandboxed terminals and messaging-app gateways.
Pilot Protocol
Open-source overlay network that gives AI agents permanent virtual addresses, encrypted UDP tunnels, NAT traversal, mutual-trust handshakes and a signature-verified app store, with native MCP setup for Claude Code, Cursor and Codex.
Moadim
A loop engine that runs coding agents on a schedule against your codebase. A loop pairs a prompt, a schedule and an agent (Claude, Codex, Hermes, NanoClaw or Pi), and each run executes in a throwaway workbench so a hung run gets killed rather than left behind. Everything is exposed twice, as a REST endpoint and as an MCP tool, so another agent can schedule work as easily as you can. Routines live in ~/.config/moadim/ and are git-trackable. It runs entirely on your own machine: no cloud, no account, no data leaving the host.
Orca
Open-source Agent Development Environment for running a fleet of AI coding agents at once. Every task gets its own git worktree, its own agent terminal, and its own browser tab, so you can fan one prompt across several agents, compare the diffs, and merge the winner. It drives roughly 30 CLI agents you already pay for, including Claude Code, Codex, OpenCode, Cursor and Copilot, and ships on macOS, Windows, Linux, iOS and Android.
Parallel Code
Free, open-source desktop app that dispatches AI coding agents in parallel, each in its own git worktree. It runs the real CLI agents you already trust (Claude Code, Codex, Gemini) side by side, then lets you review the diffs, merge the wins, and toss the rest.
Browse vetted vibe coding agencies for MVP builds, cleanups, and audits.
ready to try kit?
go to the official site and judge Kit yourself.
try kit now