Show HN: Hoop – A sandboxed P2P live collaboration harness for Claude Code

I'll pull the GitHub README so the paragraphs stick to real product details and don't invent numbers or claims.bruno-de-queiroz posted Hoop on Hacker News as a Show HN, pointing at the public GitHub repo bruno-de-queiroz/hoop. The listing frames Hoop as a sandboxed P2P live collaboration harness for Claude Code. At capture it had 1 point and 0 comments on item 48999928, so the product is still at first public signal rather than a ranked discussion thread.

Hoop is not a reimplementation of Claude Code. It is a harness around it: an install path that ships a setup wizard and a containerized local web dashboard. Runtime splits into two containers—an agent-sandbox that owns the claude binary, its own Claude OAuth profile, plugins, MCP config, sessions, and events.db, and a dashboard container that holds no credentials and never runs claude. The trusted side exposes a small HTTP API only over a Unix domain socket; the Next.js UI on localhost proxies through that socket. Live observability is push-based: hooks POST events to /ingest, the sandbox writes SQLite with FTS5 (and optional semantic search via sqlite-vec), and the dashboard tails via Server-Sent Events. Pairing for multi-person co-drive uses a cloudflared tunnel baked into the dashboard image, with share links and per-peer tokens rather than opening the agent process to the public internet.

For engineers who already run Claude Code as a daily agent, the useful claim is multi-person control without merging credentials or trusting a shared host shell. Peers can watch the same live transcript, chat in-thread, and co-drive the model; /plan forces a read-only investigation, then a review panel with inline comments, Approve, or Request changes before implementation. That maps to real workflows—pair debugging, plan review from a phone, and keeping auto-runnable MCPs and skills inside a profile that survives rebuilds—without requiring every participant to install Claude Code or Node on the host. Standalone mode is Docker-first: the sandbox seeds its own claude binary and account, so a bare machine with Docker, git, and bash can run the loop.

Market context is the crowded Claude Code extension layer: memory plugins, code-graph RAG, platform MCPs, skill packs, and ad hoc pairing scripts. Hoop’s positioning is curation plus isolation plus live co-presence: the wizard installs third-party MCPs and skills rather than re-coding them, and the security model treats the OS process boundary as the trust boundary so a compromised web UI cannot read host Claude credentials. That is closer to a sandboxed agent control plane than to another chat UI. Competitors in the same HN Claude/Codex/Fable bucket tend to ship either better single-user tooling or remote agent runners; few combine localhost-only single-user defaults with optional P2P pairing and plan-approval gates in one install.

Practical next steps for builders: if you already use Claude Code, the in-product path is marketplace add for bruno-de-queiroz/hoop, install the hoop plugin, then /hoop:setup and /hoop:dashboard. If you want isolation from the host toolchain, clone the repo, run hoop start, hoop setup, hoop login, and open the dashboard on localhost. Watch whether pairing holds under real multi-peer load, whether plan-mode permission gating stays strict as tool surfaces grow, and whether the roadmap items—injecting skill triggers into existing sessions, per-skill ephemeral containers, and non-Claude clients such as Cursor or Codex where MCPs overlap—ship before the harness becomes another abandoned Claude plugin scaffold.