1jehuang/jcode

What it solves

Jcode is a high-performance coding agent harness designed to increase the "skill ceiling" of AI coding assistants. It addresses the resource inefficiency and slow startup times common in many CLI-based coding agents, while providing a sophisticated memory system and multi-agent collaboration tools for complex, multi-session workflows.

How it works

Jcode operates as a server-client architecture that supports a wide variety of LLM providers via OAuth or API keys. It implements a semantic memory system where every conversation turn is embedded as a vector and stored in a memory graph. A "memory sideagent" handles the extraction and verification of relevant memories to inject them into the conversation without consuming excessive tokens.

For collaboration, it uses a "Swarm" system that allows multiple agents to work in the same repository. The server manages these agents, notifying them of file changes made by others and providing messaging channels (DMs and broadcasts) to resolve conflicts and coordinate tasks. Agents can also autonomously spawn their own worker agents to parallelize work.

Who it’s for

Developers who need a resource-efficient, highly customizable coding agent that can handle large-scale projects across multiple sessions and collaborate with other AI agents.

Highlights

• Extreme Performance: Optimized for low RAM usage and near-instant boot times (14ms time to first frame).
• Semantic Memory: A passive, human-like memory system that uses cosine similarity and a sideagent for automatic information recall.
• Agent Swarms: Native multi-agent collaboration with automatic conflict resolution and the ability for agents to spawn their own teammates.
• Rich TUI: Includes a side panel for real-time file updates, diff viewing, and high-speed Mermaid diagram rendering.
• Broad Provider Support: Integrates with Claude, OpenAI, Gemini, GitHub Copilot, Ollama, LM Studio, and various OpenAI-compatible endpoints.