1jehuang/jcode

jcode – a lightweight, high‑performance AI‑agent harness

What it is – jcode is a terminal‑based development assistant that runs locally (or against any OpenAI‑compatible LLM endpoint). It provides a chat‑style UI with side panels, live diagram rendering, and a built‑in memory graph so the agent can recall earlier turns without extra tool calls. The project also includes a “swarm” mode that lets multiple agents collaborate on the same repository, automatically handling file‑change notifications and conflict resolution.

Key capabilities

Feature	What you get
RAM‑efficient runtime	Baseline memory use ~28 MiB (no local embeddings) and ~167 MiB with embeddings – up to 13× less than competing CLI agents.
Fast startup	First UI frame appears in ~14 ms; first input ready in ~49 ms, far quicker than other tools.
Semantic memory	Every turn is embedded into a vector and stored in a graph. The system continuously extracts, consolidates, and queries memories, giving the agent “human‑like” recall without extra token cost.
Side‑panel UI	Load files, view diffs, or display Mermaid diagrams inline. A custom mermaid‑renderer (mermaid‑rs‑renderer) makes diagram drawing ~1800× faster than browser‑based solutions.
Swarm collaboration	Spawn any number of agents in the same repo; they auto‑notify each other of file edits, can DM or broadcast, and can be orchestrated into coordinator‑worker teams.
Broad provider support	Built‑in OAuth flows for Claude, OpenAI, Gemini, Copilot, Azure, Alibaba, Fireworks, Novita, MiniMax, Meta‑Muse, LM Studio, Ollama, plus generic OpenAI‑compatible endpoints (including self‑hosted vLLM).
Custom terminal	Uses the handterm terminal implementation for smooth scrolling and >1000 fps rendering, eliminating flicker.
Extensible config	Provider profiles, extra request‑body fields, per‑model context windows, and streaming timeout tweaks are all configurable via ~/.jcode/config.toml or CLI flags.

Installation

# macOS / Linux (one‑liner installer)
curl -fsSL https://jcode.sh/install | bash

# Windows 11 PowerShell
irm https://jcode.sh/install.ps1 | iex

For Homebrew, source builds, or custom provider setup see the “Detailed installation” section of the README.

Quick start

1. Log in to a model provider, e.g. jcode login --provider openai (or any of the supported providers).
2. Run a chat: jcode run "Write a Python function that parses CSV files".
3. Use side‑panel commands like /load path/to/file.py or /mermaid "graph LR; A-->B" to view content without leaving the chat.
4. To enable the memory side‑agent, add memory = true in the config or start the session with jcode --memory run ….

Why it matters

• Speed & cost – By keeping the process in memory and avoiding repeated embedding calls, jcode reduces both RAM pressure and token usage, which is crucial when running many concurrent sessions (the benchmarks show modest growth of only ~10 MiB per extra session).
• Collaboration – Swarm mode turns a single LLM into a multi‑agent team, letting developers parallelize tasks such as code generation, review, and testing within the same repo.
• Local‑first – All UI rendering happens in the terminal; no browser is required, making it suitable for low‑resource environments or remote SSH sessions.

Further reading – The project links to a website (jcode.sh) with full docs, SDK reference, and benchmark tables, plus a Discord community for support.

---

All details above are taken directly from the repository’s README; no additional features have been inferred.