jcode: a Coding Agent Harness in Rust That Starts in 14ms
jcode is a coding agent harness written in Rust, and it is on GitHub Trending with 8.8k stars, up 199 today. Version 0.52.0 shipped yesterday, July 19 — release number 125-plus, which tells you the cadence. MIT licensed, by 1jehuang.
The pitch is performance as a spec sheet: 14 milliseconds to first frame, which the README claims is 100x faster than competitors, and 27.8 MB of memory for a single session. The README benchmarks it directly against Claude Code, Codex CLI, Cursor Agent, Copilot CLI and OpenCode. Harness competition has reached the stage where startup time and RSS get published like GPU specs.
Under the hood there is more than speed. Semantic memory stores extracted facts as vector embeddings in a graph, recalled by cosine similarity instead of burning context tokens. Swarm mode runs multiple agents in the same repo with automatic conflict management and inter-agent messaging. And self-dev mode lets the agent modify its own source, rebuild and reload — the self-improving-harness pattern shipping as a checkbox feature. Providers: Claude, OpenAI, Copilot, Gemini, plus local models via Ollama and LM Studio. MCP servers and a Firefox browser bridge included.
Repo: https://github.com/1jehuang/jcode
← Back to all articles
The pitch is performance as a spec sheet: 14 milliseconds to first frame, which the README claims is 100x faster than competitors, and 27.8 MB of memory for a single session. The README benchmarks it directly against Claude Code, Codex CLI, Cursor Agent, Copilot CLI and OpenCode. Harness competition has reached the stage where startup time and RSS get published like GPU specs.
Under the hood there is more than speed. Semantic memory stores extracted facts as vector embeddings in a graph, recalled by cosine similarity instead of burning context tokens. Swarm mode runs multiple agents in the same repo with automatic conflict management and inter-agent messaging. And self-dev mode lets the agent modify its own source, rebuild and reload — the self-improving-harness pattern shipping as a checkbox feature. Providers: Claude, OpenAI, Copilot, Gemini, plus local models via Ollama and LM Studio. MCP servers and a Firefox browser bridge included.
Repo: https://github.com/1jehuang/jcode
Comments