27 MB idle.

A coding agent in one Rust binary. Bring your Claude or ChatGPT login and start.

Memory when idle
agent27 MB
jcode48 MB
pi310 MB
2.30 sto answer "OK" pi takes 4.20 s
17.8ktokens of base context jcode sends 24.2k
$0.61for the semver port jcode costs $0.84
9.5tool calls per task jcode needs 14.7

Small binary, full agent

Benchmark: same model, same prompt, 6 runs each, every run passed all hidden tests. Full table

Every account you have

Claude, ChatGPT and GitHub Copilot logins, Anthropic and OpenRouter keys. When one hits its limit, it moves to the next. /usage shows what is left.

Local and open models

Run offline on Ollama or LM Studio, or bring a key for one of 38 hosted providers, from DeepSeek and Groq to Kimi and Z.AI. Any OpenAI- or Anthropic-compatible server works with one line of settings.

Edits by line anchor

If the file changed since it last read it, the edit is refused. No silent overwrites.

Checks its own work

Language server errors and your edit hooks after every change. It fixes what they find.

Rewind

Esc twice goes back to an earlier prompt, and your files go back with it.

Subagents

Big work splits into parallel child agents, tuned in a plain Markdown prompt.

Plugins, MCP, skills

Plugins in any language over JSON-RPC. MCP servers with OAuth. Skills, with /skill-usage to find the dead ones.

Scripts and CI

agent -p prints and exits, --headless streams JSONL, agent acp talks to Zed.

One line to install.