// the find
tiann/hapi
App for Claude Code / Codex / Gemini / OpenCode, vibe coding anytime, anywhere
HAPI wraps AI coding agents (Claude Code, Codex, Gemini CLI, OpenCode) running on your local machine and exposes them to a web UI, PWA, or phone so you can approve actions, send messages, and monitor progress remotely. It's a remote control layer, not a replacement — the actual agent runs on your dev box and HAPI handles the communication bridge. Aimed at developers who want to kick off long-running agent tasks and then step away.
The relay uses WireGuard + TLS for E2E encryption, which is the right call for a tool that punches a hole to your dev machine — not bolted on later. Test coverage is real: .test.ts files throughout including protocol fixtures for ACP message parsing, not just happy-path smoke tests. The agent abstraction layer is clean — separate directories per agent (claude, codex, cursor, kimi, opencode) each with their own launchers and protocol adapters, so adding a new backend doesn't require touching everything. Self-hosted relay options via Cloudflare Tunnel or Tailscale are first-class, not an afterthought.
The managed relay is the path of least resistance, but it means a third party is involved in tunneling to a machine where an AI agent can execute arbitrary shell commands — the encryption claim doesn't change the threat model if you don't control the relay. The `npx @twsxtd/hapi` install pattern with a weirdly namespaced package (`twsxtd` instead of a recognizable org) is a red flag for a security-sensitive tool; you're running arbitrary code on first invocation. The multi-agent support is genuinely complex — ACP backend, Claude SDK mode, Codex app-server mode, Cursor legacy vs ACP — and the maintenance burden of keeping all these working as upstream agents evolve their protocols will be significant. Windows support appears absent from the CLI binary distribution (only darwin-arm64, darwin-x64, linux-arm64, linux-x64 in `cli/npm/`).