agent-harness

agent-harness runs the same lifecycle regardless of which host you pick: discover, recommend, mirror, stage, activate, then wire. That consistency matters because the underlying tools don't agree on anything. A skill that works in Claude Code isn't packaged the same way as one for Cursor, and this tool smooths out those differences behind a single command surface. One surface, one lifecycle.

The workspace <host> command is the default path for new users. It runs broad discovery, ranks the candidates, stages and activates the chosen assets, and finishes with the host's wire-in, all from a single line like agent-harness workspace cursor --intent frontend. Intents such as devops, design, or research steer the selection toward your kind of work, and you can pass more than one. Setup is one line. That's the appeal.

Wiring a host is non-mutating unless you ask otherwise. By default, wire <host> writes a reviewable plan and touches nothing. You can read the exact target paths and planned writes, then apply or reset when you're ready. For anyone who's had a tool quietly rewrite editor settings, that's a welcome line in the sand.

Rather than pulling assets straight from the internet into your workspace, the tool mirrors a bounded set into reproducible local artifacts and pins each generation. Candidates that look risky get routed through a quarantine folder instead of being activated. The result is a setup you can inspect and roll back rather than one that shifts under you.

agent-harness implements the Agentic Resource Discovery 1.0 spec, the standard backed by Google, Microsoft, Hugging Face, GitHub, and NVIDIA. It publishes a .well-known/ai-catalog.json so registries can index its curated assets, and it can consume other ARD registries as discovery sources. Its tagline sums up the intent: discover anywhere, install everywhere.

The project registers adapters for VS Code with Copilot, OpenCode, Cursor, Zed, Claude Code, Pi, and OpenAI Codex. Each adapter owns its host's files, settings, apply and reset behavior, and safety boundaries. Some reuse another host's package layout while ranking through their own policy, which keeps Cursor, for example, compatible with Copilot-style assets.

Recommendations weigh policy, demand signals, trust, cost, and diversity, not just popularity. Official and verified sources are preferred, and first-party sources get demoted when the owner evidence doesn't check out. The upshot is fewer sketchy packages landing in your workspace and more of the ones that actually fit what you're building.