What it does
Keen Code is a terminal-based AI coding agent in the vein of Claude Code or Codex CLI, written in Go. The README calls it opinionated and minimalistic: it keeps the agent harness simple and ships ten built-in tools for reading, writing and editing files, searching, running shell commands, fetching web pages, asking the user questions, delegating tasks and calling MCP tools. You switch between providers with /model, including Anthropic, OpenAI, Codex through OAuth, Gemini, DeepSeek, Kimi, GLM, MiniMax, OpenCode Go, Yolo-Auto and Amazon Bedrock. It supports MCP servers over streamable HTTP or stdio, skills activated as slash commands, subagent profiles defined as markdown files, persistent memory files, context compaction, and a headless mode started with keen run. Edits use hash-anchored line references, so stale edits are rejected rather than applied to changed content. A permission system checks file access, and a separately configured second model can act as an adversarial reviewer. You install it with a shell script or npm. The README also says the code was written by AI agents under a human engineer's direction, as an experiment.
Capabilities
- Switch between many model providers with /model
- Hash-anchored file edits that reject stale anchors
- MCP servers, skills and parallel subagents
- Permission system for sensitive file paths
- Headless mode for non-interactive runs
- Optional adversarial review by a second model
When to use it
- You want a lightweight terminal coding agent that you can point at your own model provider
- You want to move between providers inside one agent
- You want to run the agent from scripts without the interactive interface
- You are curious how a coding agent built largely by coding agents works
