A zero-setup local web app to find, preview, and copy your AI-agent assets — skills, agents, rules, MCP servers, instructions, and plugins — across tools (Claude Code, Codex, Cursor, Windsurf, …).

npx agent-shelf

That launches a Go loopback server (bound to 127.0.0.1, session-cookie + loopback-origin gated) which embeds a React SPA and scans your home + known tool directories. The browser opens a three-column control room: a tool/kind/scope sidebar, a virtualized searchable library, and an inspector that previews the file and copies it into another tool or project with a live destination preview, inline conflict resolution, and a 10s undo.

Everything runs on your machine — no network calls, no telemetry.

make dev          # prints the two dev processes to run
cd web && pnpm dev                          # SPA + HMR on http://127.0.0.1:5173
cd server && go run ./cmd/agent-shelf serve # Go server; prints a http://127.0.0.1:<port>/?token=… URL

The SPA's API calls only work against the Go server (the embedded build), so for end-to-end behavior use make build then run the binary; pnpm dev alone renders the shell with no backend.

make test         # go test ./...  +  vitest (web unit)
make e2e          # Playwright (spins real Go servers + the Vite dev shell)
make bench-budget # scan/search/first-paint benchmarks

make build        # SPA -> embed -> cross-compile all platform binaries + resync shasums

Publishing to npm is documented in AGENTS.md → Release.

Read AGENTS.md first — it has the architecture, the API/data model, the exact build/test/release commands, and the non-obvious gotchas (the embed-at-compile model, full-library search, npm publish quirks, …). Design spec lives in specs/001-web-mvp/.