Zendesk for StartupsUp to 2 years free for startups + AI Agents & Copilot
Promoted
Launching today
agent-harness
The missing package manager for MCP servers and AI skills
1 follower
The missing package manager for MCP servers and AI skills
1 follower
Stop managing AI agent assets by hand. agent-harness discovers MCP servers, skills, and rules from trusted sources; pins and mirrors them reproducibly; quarantines risky inputs; then wires everything into Cursor, Claude Code, Copilot, Zed, Codex and Pi.
Hey PH š ā I'm Ahmed, the solo developer behind agent-harness.
The problem I kept hitting: every AI editor has its own file layout, its own rules format, and its own way of consuming MCP servers and skills. I'd wire up Cursor, then rebuild everything from scratch for Claude Code, then again for Copilot. And not to mention the differences in agent assets required based on the project stack and session intent. There was no reproducible middle layer ā just scattered files and hope.
agent-harness is that middle layer. One CLI that discovers agent assets from trusted sources, pins and mirrors them locally, quarantines anything that looks like a prompt-injection risk, and wires the result into whichever editor you're in ā Cursor, Claude Code, VS Code/Copilot, Zed, OpenCode, Pi, or Codex.
Three things I built that I haven't seen elsewhere:
ā A quarantine + audit trail for risky assets (no silent installs)
ā Preview-first wire-in with explicit apply/reset ā no mystery files
ā ARD-compatible (v1.0 spec, announced by Google + Microsoft) as both publisher and consumer
v2.1.0 ships today on npm:
npm install -g @ar27111994/agent-harness
agent-harness workspace opencode --intent general
Full walkthrough: https://youtu.be/u1OmcS97iOg
Happy to answer any questions ā especially about host-specific wiring, the quarantine model, or ARD integration plans for > v2.1. What editor are you running?