Overview of jcode: A Rust-Based AI Coding Agent Harness

Performance and Architecture

jcode is an open-source AI coding agent harness written entirely in Rust, designed to minimize resource overhead compared to existing agent frameworks. It operates with a baseline memory footprint of approximately 28 MB and boots in 14 milliseconds. Each additional concurrent session adds roughly 10 MB of RAM usage, allowing for efficient parallel agent execution without the high resource costs associated with other tools.

Core Features

• Agent Memory: The system utilizes vector embeddings and semantic search to maintain context across sessions. Memory retrieval occurs passively, injecting relevant historical context into the conversation without requiring explicit tool calls.
• Swarm Mode: Multiple agents can operate within the same repository simultaneously. The system manages file-access conflicts automatically and allows agents to communicate with each other to coordinate tasks.
• Browser Automation: A built-in Firefox backend enables agents to perform web-based tasks, including clicking elements, typing, taking screenshots, and evaluating JavaScript, without requiring external MCP servers.
• Self-Modification: The agent supports a self-dev mode, allowing it to modify its own source code, rebuild, and reload during a session.
• Compatibility: It supports a wide range of providers, including Claude, OpenAI, Gemini, GitHub Copilot, Azure, OpenRouter, DeepSeek, Ollama, and LM Studio. It can also resume sessions initiated in other harnesses like Claude Code or OpenCode.

Usage

Users can install the tool via a shell script on Linux or macOS, or via Homebrew. The interface provides a custom terminal-based TUI that renders diffs and Mermaid diagrams in real-time. For non-interactive workflows, the tool supports a command-line execution mode (jcode run) and a background server mode (jcode serve).