Jcode — AI Coding Agent

Jcode is an open-source (MIT) AI coding agent for the terminal, written in Rust by Solo Systems (backed by Combinator). It's built around three bets: massive parallelism, harness-over-model, and fully open dev tools.

The result is the most RAM-efficient agent on the market — about 10.4 MB per additional session (20× less than Claude Code) and a 48.7 ms time-to-first-input (72× faster than Claude Code). Run dozens of agents in parallel on consumer hardware.

Navigation

• Getting Started — Install, provider login, first session
• Prompt Engineering — Hill-climbable goals, confidence stepping, AGENTS.md conventions
• Configuration Reference — config.toml, MCP, memory, swarms, remote gateway
• Memory, Swarms & Self-Dev — Durable memory, multi-agent swarms, background tasks, self-dev mode

Quick Start

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

# macOS via Homebrew
brew tap 1jehuang/jcode
brew install jcode

# Windows 11 (PowerShell 5.1+)
irm https://jcode.sh/install.ps1 | iex

Log in with a provider you already pay for, then start a session:

jcode login --provider claude     # Claude subscription
jcode login --provider openai     # OpenAI / ChatGPT / Codex
jcode login --provider gemini     # Google Gemini
jcode login --provider ollama     # local Ollama server

cd my-project
jcode

What Makes Jcode Different

• RAM efficiency — ~10.4 MB extra per session vs ~213 MB for Claude Code. Parallel agents without a server farm.
• Memory — every turn is embedded as a semantic vector; related memories are recalled automatically from a graph via cosine similarity.
• Swarms — coordinated multi-agent teams with file-conflict detection and model routing via .jcode/swarm-prompt.md.
• Self-dev mode — tell jcode to enter self-dev mode and it edits, builds, tests, and reloads its own binary.
• MCP that never blocks startup — MCP tools are advertised from an on-disk schema cache, so the prompt cache stays warm.
• Background tasks — run_in_background tasks survive reloads, render live progress cards, and wake the agent on events.
• Append-only context — a stable prompt prefix keeps the provider KV cache hot, cutting cost and latency.

Core Workflows

Run a single prompt non-interactively, resume by name, or serve persistently:

Project instructions live in AGENTS.md at the repo root; machine-wide instructions in ~/AGENTS.md. Both load into every session in scope.

Frequently Asked Questions

Related Resources

• Getting Started — install, configure, and begin your first session
• Prompt Engineering — hill-climbable goals, confidence stepping, todo discipline
• Configuration Reference — config.toml, MCP, memory, swarms, remote gateway
• Memory, Swarms & Self-Dev — advanced workflows
• Tool Comparison — jcode vs opencode vs pi vs Claude Code