Open-source jcode claims 14 ms startup and multi-agent repo coordination
thisguyknowsai · x · 2026-07-27
A developer has open-sourced jcode, a coding-agent harness that claims much faster startup and much lower memory use than Claude Code.
What it says jcode does
- First frame renders in 14 ms; the post compares that with 3,436 ms for Claude Code.
- One active session uses 27.8 MB RAM versus 386.6 MB for Claude Code.
- Ten parallel sessions stay at 117 MB, while OpenCode reportedly grows to 3.2 GB.
Core design
- Each agent gets a semantic memory graph instead of a scratchpad.
- Every turn is embedded, queried for related memories, and verified by a side agent before being injected back into context.
- Background consolidation checks for stale or conflicting facts, avoiding manual /remember calls and token burn on lookup tools.
Multi-agent workflow
- The Swarm system lets two agents work in the same repo.
- If one edits a file another is reading, the other is pinged and can inspect the diff.
- Agents can DM each other, broadcast to the room, or spawn worker teams for parallel tasks.
Other details
- Supports 30+ providers including Claude, ChatGPT, Gemini, GitHub Copilot, Azure, OpenRouter, DeepSeek, Groq, Mistral, Perplexity, Fireworks, Ollama, and LM Studio.
- Can resume broken sessions from tools like Claude Code, Codex, OpenCode, or pi.
- Written in Rust, licensed MIT, works on macOS, Windows, Linux, and Termux, and has a native iOS app coming.
More from coding & agent
- Memoria open-sources Git-style version control for AI agent memory — bigaiguy · 2026-07-27
- New ComfyUI prompt manager node adds on-the-fly edits, randomization, and shared presets — 3deal · 2026-07-27
- VS Code Remote makes the Claude-and-Codex server setup usable anywhere — nico_jeannen · 2026-07-27
- Developer runs Claude and Codex on a server to work from a phone or iPad — nico_jeannen · 2026-07-27
- Tanuki Context claims up to 94% token savings by turning some LLM inputs into images — 0syna · 2026-07-27
- MyClaw expands into multi-agent hosting with OpenClaw and Hermes — techNmak · 2026-07-27