jcode Open source

What it does

jcode is a terminal AI coding agent harness built in Rust, whose headline claim is very low memory usage (the README reports about 27.8 MB RAM for a single session with embeddings off). It supports multi-session, multi-agent collaborative workflows against the same repository.

What is inside

• A semantic memory system using vector embeddings for automatic context retrieval
• "Swarm" mode letting multiple agents work in the same repo with automatic conflict resolution
• Built-in Firefox-based browser automation for web-interaction tasks
• A self-development mode where agents can modify jcode's own source, rebuild and reload it
• Broad model support: Claude, OpenAI, GitHub Copilot, Gemini, Azure and other providers

Works with

Designed as a provider-agnostic harness, with documented support for Claude, OpenAI, GitHub Copilot, Gemini and Azure model backends.

How to install or connect

curl -fsSL https://jcode.sh/install | bash

Windows:

irm https://jcode.sh/install.ps1 | iex

Maintenance and safety

MIT licensed, with detailed documentation, benchmarks and an active Discord; the repo shows roughly 9,400 commits and regular updates. That said, this is largely a solo/small-team project relative to its very high star count, and the self-development feature — letting agents rebuild and reload their own source automatically — is a meaningfully higher-risk capability than a normal file-editing agent. Anyone enabling it should sandbox the agent and review changes before they take effect, and should independently verify the RAM and performance benchmarks rather than taking them at face value.

Who should use it

Developers on memory-constrained machines who want a lightweight multi-agent "swarm" harness with semantic memory, and who are comfortable auditing an agent that can, if configured to, modify and reload its own code. It's a reasonable pick for hobbyists or small teams running many concurrent agent sessions on modest hardware, but larger organizations should pilot it carefully, pin dependencies, and disable the self-development mode unless they specifically need it and can review the changes it produces.