Airnode B

Bifrost — talk to your coding agent from your phone

A voice-controlled opencode: speak or type on your phone (PWA), a LiveKit voice agent bridges your words to a local opencode server, and you get the answer back as text and speech — with markdown, attachments, multi-session support, and frontend diagnostics.

Easiest install: let an agent do it (Omarchy, Arch, Debian, any Linux box)

Open your coding agent (opencode, Claude Code, Codex, …) in an empty folder and give it this one line:

https://raw.githubusercontent.com/veraldar/bifrost/v0.6.0/skills/install/SKILL.md set it up

That's the whole instruction. The agent fetches the skill, clones this repo, checks prerequisites (Node ≥ 22, docker for voice), mints every secret, starts LiveKit + opencode + the voice agent + the PWA, verifies each layer, and prints a verdict + the URL to open. Already cloned? Say instead: "follow skills/install/SKILL.md and set it up". One manual step remains afterwards: opencode auth login (your LLM provider key — nothing may mint that for you). Prefer manual setup? Keep reading.

One-click Vercel deploy of the PWA: after cloning, set the env vars LIVEKIT_URL, LIVEKIT_API_KEY, LIVEKIT_API_SECRET (see pwa/.env.example). Text-mode needs OPENCODE_URL reachable from the deployment — keep opencode private (Tailscale), it has shell access!

Using other AIs (Claude Code, Codex, …)

bifrost drives opencode natively — sessions, voice, queue, transcript. Other AI CLIs are one sentence away, not one integration: from any session, just ask — "ask claude to review this diff" or "have codex write tests for X" — and the delegate skill runs them as sub-tools, relays their output into your transcript, and works by voice too. No config, no second login flow beyond the CLI's own auth.

Your voice, your identity

Bifrost is built to run speech on your own hardware — your voice never has to leave your machines. That's the setup we push. But you choose: local models (CPU or MLX), or any cloud API for LLM and voice — American, Chinese, European, open weights. The AI is yours to pick; the identity stays yours.

Architecture

Hardware — 3 nodes, one Tailscale tailnet (no public exposure):

Software layers (top = user, bottom = model):

Phone (PWA: Next.js)                        Omarchy desktop
┌──────────────────────────┐                ┌──────────────────────────────────┐
│ sessions UI / transcript │   REST         │  opencode serve :4096            │
│ text + attachments  ─────┼───────────────►│  (sessions, prompts, streaming)  │
│                          │                │        ▲                         │
│ WebRTC media (mic/speech)│                │        │ llm.LLM adapter         │
│   ▼                      │                │  Python voice agent (livekit-    │
│ LiveKit server :7880 ◄───┼───WebRTC───────┤  agents)                         │
│ (docker, host net)       │                │   STT ⇄ opencode ⇄ TTS           │
└──────────────────────────┘                └───────┬─────────────┬───────────┘
        ▲                                           │ LAN         │ fallback
   tailscale serve                         Mac Studio :8001    speaches :8000
   (tailnet-only TLS)                      (MLX, GPU)          (docker, CPU)

Local models (no cloud speech APIs):

Everything is systemd-managed on omarchy (lk-pwa, lk-agent, opencode-serve) + docker compose — reboots self-heal.

• PWA (pwa/) — Next.js 15 + Tailwind 4. Sessions list, transcript with markdown (Streamdown), image/file attachments, sandboxed HTML preview, pixel-art UI (pixelarticons), push-to-talk & hands-free voice modes.
• Proxy (pwa/app/api/*) — the phone never talks to opencode directly; the Next server proxies sessions/messages/tokens and stays the single auth/seam point. Also receives frontend diagnostics (/api/diag), log-only.
• Agent (agent/) — Python livekit-agents worker. Custom llm.LLM adapter maps LiveKit room → opencode session (room name = session slug), so voice turns land in the same transcript as text.
• Deploy (deploy/) — docker compose: LiveKit (host network, ICE pinned to tailnet + LAN IPs so phone media flows over Tailscale) + speaches STT/TTS.
• Diagnostics — the PWA records console errors, crashes, network timings, and taps, uploads them in batches, and they are logged server-side (journalctl + pwa/.diag/*.log). Log-only by design: no automatic fix sessions (a past auto-autofix experiment was removed as a bad idea).

One-click install (self-host)

Minimum hardware: ONE Linux box (any machine that runs Docker + Node) + your phone. No Mac needed — the CPU speech fallback (speaches, in the compose) works out of the box; a Mac Studio running MLX speech models is an optional quality upgrade (see agent/.env.example → SPEACHES_URL). Everything below runs on that one box; nothing needs to be public — Tailscale is the only network requirement (install it on box + phone).

git clone https://github.com/veraldar/bifrost.git && cd bifrost
scripts/selfhost-check.sh   # one-command check: can this box run the PWA? (clone→build→boot→probe)

# 0. opencode itself (the coding agent this whole thing drives)
#    install: curl -fsSL https://opencode.ai/install | bash   (or your package manager)
opencode auth login                          # configure your LLM provider key
opencode serve --port 4096 &                 # REST server on 127.0.0.1:4096

# 1. LiveKit + speech stack (cp example → real; edit .env LIVEKIT_KEYS and
#    livekit.yaml ICE IPs for your network — see deploy/livekit.yaml.example;
#    `tailscale ip` on the box gives the tailnet IP to pin)
cd deploy && cp livekit.yaml.example livekit.yaml && docker compose up -d && cd ..

# 2. Voice agent
cd agent && uv sync && cp .env.example .env  # edit values (SPEACHES_URL empty = CPU fallback)
uv run agent.py dev                          # or a systemd unit

# 3. PWA
cd ../pwa && npm install && cp .env.example .env.local  # edit values
npx web-push generate-vapid-keys --json      # paste into .env.local for push notifications
npm run build && npm start                   # :8080
tailscale serve --bg 8080                    # tailnet-only HTTPS

Or deploy just the PWA to Vercel with the button above and point it at your LiveKit (and a network-reachable opencode).

Env vars

Roadmap

One structural change per version; each is gated (scripts/gate.py) and lives on daily use before the next. Every version is a frozen release you can switch back to in about a minute (scripts/promote.sh --list / --switch / --rollback). Full contract: docs/STABLE.md.

docs/plan.md keeps the working board; docs/environments.md records installs and test runs.

Support

Want to be part of the journey? Veraldar is free and stays free — but it isn't free to build.

• GitHub Sponsors: github.com/sponsors/veraldar (pending activation)
• OpenCollective: opencode Collective pending — transparent public ledger

Every franc goes to hardware, hosting, and keeping the voice local.

License

Bifrost is AGPL-3.0 — copyleft including network use: modified versions offered as a service must make their source available. Talking to bifrost over its API does not make your app a derivative; interop stays free.

Attributions: PWA derives from livekit-examples/agent-starter-react (Apache-2.0). Icons: pixelarticons (MIT).