Introduction

jcode is a terminal-based coding agent harness written in Rust. Instead of being another chat UI, it targets heavy multi-agent workflows: launching many sessions in the same repository without exhausting memory or making the interface laggy. It runs as an interactive CLI/TUI, supports Claude and OpenAI Codex OAuth among other providers (provider list), and ships its own collaboration layer for agents.

The core mechanism behind parallel work is "Swarm". When two or more agents are alive in the same repo, a local server coordinates them: if agent A edits a file that agent B has already read, agent B receives a notice and can inspect the diff and resolve conflicts. Agents can also message each other directly, broadcast to every agent on the server, or broadcast only to agents in the same repo (Swarm docs). An agent can create sub-agents itself, becoming a coordinator while the spawned agents act as workers; groups, channels and completion status are managed automatically, in headed or headless mode.

A second distinctive feature is self-development: the agent can be told to enter self-dev mode, edit jcode's own source, build and test it, then reload the binary and continue in the existing sessions (self-dev section). Token optimization, KV cache handling and hill-climbable measurements are documented in the harness notes.

It suits developers who already delegate work to coding agents and want to fan out several of them over one codebase, rather than users looking for a graphical assistant. The interactive screenshots, including inline images, a session resume menu, inline LaTeX and a cache-cold warning, are collected on jcode.sh.

Quick Start

Prerequisites: macOS or Linux, a terminal, and credentials for at least one supported model provider (for example Claude or OpenAI Codex OAuth).

1. Install with the official script:

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

2. Launch jcode in the repository you want it to work on, then configure or authenticate a provider when prompted; the provider documentation is linked from the docs.
3. Start a session and give the agent a task in the TUI. Text should appear essentially immediately; the README measures 19.1 ms time to first input versus 596 ms for pi and 3.5 s for Claude Code in its own benchmark (benchmark table).
4. To try parallel work, start a second session in the same repository. The server should register both agents and relay file-change notices between them.

The README does not spell out the exact prompt sequence for authenticating each provider or the precise keystrokes for enabling Swarm, so use the provider docs and the on-screen prompts rather than copying assumed commands.

Usage Constraints

• Platform: the install command is documented only for macOS and Linux; no Windows install path is given (installation).
• Requires external model access. jcode is a harness, not a model; you need an account or API access for a supported provider, and OAuth-based login is explicitly supported for Claude and OpenAI Codex.
• Swarm behavior depends on agents running against the same local server and repository; the conflict notices and messaging apply within that setup.
• Self-dev mode lets the agent modify jcode's own source and reload the binary. Expect it to touch the installed program and its build artifacts, so run it where you can tolerate that, such as a source checkout.
• The project is MIT licensed. Benchmarks in the README are project-run; they describe that specific test setup (headless sessions for RAM, ten interactive launches for first-input time) rather than guaranteed results on your machine.

HotRepo's take

jcode is one of the faster-growing repositories HotRepo has recorded in its daily snapshots. Stars rose from 8,729 on 2026-07-19 to 14,789 on 2026-08-01, with daily gains of several hundred stars in that stretch, and the count stood at 21,383 by 2026-11-01. On that November date it appeared on the Rust language daily list at position 9 with 29 stars gained that day, a much slower pace than the July–August window. The README's performance claims are central to its pitch: the project's own tables put additional-session memory at about 3.3 MB versus 169 MB for Claude Code and 194 MB for Codex CLI, and time to first input at 19.1 ms. Those numbers come from project-maintained benchmarks, so they are useful as a description of what the authors optimized for, not as an independent evaluation of response quality or reliability. The most interesting technical bet is the server-coordinated Swarm and the self-dev loop, which turn the agent into something that manages other agents and rebuilds its own harness.