jcode — RAM-Efficient AI Coding Agent for the Terminal

jcode is an open-source AI coding agent for the terminal, written in Rust — with dramatically better RAM efficiency and boot speed than Claude Code, Cursor Agent, or GitHub Copilot CLI.

One command to install:

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

Why jcode?#

RAM — 14x Lighter#

Benchmark with 1 active session:

At 10 concurrent sessions the gap widens — OpenCode uses 3.2GB (27.7×), Claude Code uses 2.3GB (19.7×), jcode stays at 260MB.

Speed — 14ms Boot#

Time to first frame (from command to visible UI):

Extra RAM Per Session#

Features#

Durable Memory#

jcode goes beyond the context window with a graph-based memory system:

• Each turn/response is embedded as a semantic vector
• Queries the memory graph via cosine similarity to find relevant past context
• A memory sideagent verifies relevance before injecting into the conversation
• Memories are extracted automatically on semantic drift, after K turns, or at session end
• Auto-consolidation reorganizes, checks for staleness, and resolves conflicts
• Explicit memory tools let the agent search or store memories on demand
• Session search provides traditional RAG over past sessions

Swarm — Multiple Agents Collaborating#

Spawn 2+ agents in the same repo — the server manages them automatically:

• When agent A edits a file agent B is reading, B receives a notification
• Agent B can ignore it or check the diff
• Messaging: DM a single agent, broadcast to all, or scope to repo agents
• Agents can autonomously spawn their own swarms — the main agent becomes a coordinator, spawned agents become workers
• Supports both headless and headed modes

UI — Side Panel, Diagrams, Info Widgets#

• Side panel: load files for real-time viewing, diff viewer, or direct agent output
• Mermaid diagrams: rendered inline using mermaid-rs-renderer ↗ — 1800× faster than JS libraries
• Info widgets: use only negative screen space, never taking room from responses

• 1000 FPS rendering — zero flicker

• Custom scrollback — far beyond native terminal scrollback
• Handterm: a custom terminal (github.com/1jehuang/handterm ↗) built for smooth, partial-line scrolling

OAuth & Providers#

Broad provider support out of the box:

jcode login --provider claude
jcode login --provider openai
jcode login --provider gemini
jcode login --provider copilot
jcode login --provider azure
jcode login --provider ollama
jcode login --provider lmstudio
jcode login --provider fireworks
jcode login --provider minimax

Plus OpenAI-compatible endpoints: OpenRouter, DeepSeek, MoonshotAI, HuggingFace, Nebius, Scaleway, vLLM — even local servers without an API key. Self-hosted vLLM instances work with a single jcode provider add command.

Custom Mermaid Renderer#

jcode ships its own Rust-based Mermaid rendering library — no JavaScript, no browser dependency:

// mermaid-rs-renderer — renders diagrams 1800× faster
let diagram = mermaid_rs::render("graph TD; A-->B;");

Config & Customization#

• Hot-reloads config.toml changes at runtime
• Every color is configurable with palette harmony scoring
• Left-aligned by default; toggle centered mode with Alt+C or /alignment
• Disable emoji globally with emoji = false in [display] config
• Per-model context window config for endpoints that don’t advertise it

Installation#

macOS & Linux#

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

Windows 11 (PowerShell)#

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

Homebrew#

brew install jcode

Quick Start#

# Run with Claude
jcode --provider claude run "explain the code in src/"

# Run with OpenAI
jcode --provider openai --model gpt-4o run "refactor this file"

# Run with local Ollama
jcode --provider ollama --model llama3.2 run "hello"

# Login to a provider
jcode login --provider claude

# Start interactive session
jcode

Comparison#

Conclusion#

jcode is an AI coding agent engineered for performance — Rust, zero-cost abstractions, a graph-based memory system, and native swarm collaboration.

Its key advantages:

• Ultra-low RAM: 27.8 MB baseline — runs on low-end machines and scales to many concurrent sessions
• Instant boot: 14ms — no waiting
• Graph memory: persists context across sessions with automatic consolidation
• Swarm: multiple agents collaborating autonomously

Try it:

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

References#

• jcode GitHub Repository ↗
• jcode Website ↗
• jcode Benchmarks ↗
• mermaid-rs-renderer — Ultra-fast Mermaid renderer ↗
• Handterm — Custom terminal for smooth scrolling ↗