remote-workstreams

Hold a natural spoken conversation with a coding agent — Claude Code or OpenAI's Codex — while it does real agentic work. Your Mac runs everything — the audio pipeline, the sessions, the state. Your iPhone is a thin browser client reached over your own tailnet. No cloud infrastructure beyond the STT and TTS APIs; nothing between your phone and your Mac but Tailscale.

The core design: every model interaction is a real, interactive agent session, living as a window in one tmux session on your Mac. Claude Code and Codex CLI are both first-class engines — either alone is a complete install, and with both you pick per conversation and per workstream. The phone and the laptop drive the same sessions — walk over, tmux attach, keep typing. Every session inherits your full CLI setup (skills, hooks, instructions, permission rules) natively, and no model API key exists anywhere — all model use rides the CLIs' own auth. A persistent conversation session talks with you; planner and injector sessions turn that conversation into workstreams — execution sessions you watch as live cards on the phone. Claude sessions launch with --remote-control, so each one also shows up in the Claude iOS app — you get a native ping when a session goes live, and a second window into any workstream. If the phone drops — call, dead spot, Safari suspending — the sessions live on in tmux; reconnect and resume mid-conversation.

From the phone

• Talk. Streamed STT with barge-in — speak over the assistant and it stops. The transcript is the chat, so tool activity and final replies render from the session's own record. A hush toggle mutes the speaker for reading in public — replies land in chat and are never synthesized, so silence costs nothing; a keyboard pill unfolds into a composer for when you can't talk at all.
• + Workstream marks the conversation since the last launch, has a planner session distill it into a stint plan, and launches an execution session on it — one button, no review step.
• Live cards, one per workstream: state-colored title (green waiting, blue waiting with subagents running, amber mid-turn, red errored or gone), subagent count, and a context meter that doubles as the Compact button. Send latest routes the newest conversation delta through an injector session into that workstream; Check in has the conversation read the workstream's transcript and answer out loud.
• Tap a card for the workstream's detail view: the live transcript as a log (rendered instantly from cache, streamed from the session's own record) over a composer that queues typed messages straight into the session — no injector, nothing forked. A message typed while the session is mid-turn shows immediately as a faint queued line, then as a full line when the session takes it. The card's Compact/context meter rides beside the keyboard pill; everything else stays on the card.
• Approvals. Destructive shell commands inside Claude workstreams relay to the phone as approve/deny cards; everything else runs without interruption.
• The picker (hamburger) sets the conversation's and future workstreams' engine and model, and a plans row picks who runs the planner and injector — opus thinking at high or luna at xhigh. Every row offers only the engines wired on the Mac — a Codex-only box shows the Codex tiers alone. Every button on the phone arms-then-confirms — the armed label states the consequence (blue swap? is safe, red clear? wipes the conversation).

Engines

The model name carries the engine: pick a Claude model and the session is Claude Code; pick a Codex model and it's Codex CLI (model lists live in remote_workstreams/engines.py). A Claude-to-Claude conversation pick switches the live session in place; switching engines (or between Codex models) starts a fresh conversation, announced before you confirm. Running workstreams always keep the engine and model they launched with. Codex workstreams are renamed to their stint title in ChatGPT as they launch and archived when you tap End. The planner and injector sessions behind + Workstream and Send latest run on whichever engine installed the service (store-configurable), so a single-engine box is fully functional either way.

Engine differences that show: Codex has no known-in-advance session id, so the service discovers each session's rollout file on disk; Codex workstreams run with full filesystem access and approval prompts disabled because the phone-approval relay is a Claude Code hook; and Codex sessions don't resume across service restarts — a dead Codex conversation window starts fresh. Sol and Terra use Codex's V2 multi-agent runtime.

The launchd service runs codex from its Homebrew-aware PATH. Codex conversations also appear in the ChatGPT app's history, so a live workstream can be opened there without separate pairing.

Topology

iPhone (Safari PWA) ──WebSocket/HTTPS over Tailscale──> Mac
                                                         ├─ FastAPI service (launchd, persistent)
                                                         │   ├─ Audio pipeline: Deepgram STT ⇄ VAD ⇄ Cartesia TTS
                                                         │   ├─ tmux session "voice": convo + workstream
                                                         │   │    sessions — Claude Code or Codex CLI
                                                         │   │    (attach from any terminal)
                                                         │   ├─ Transcript tailing: the session's JSONL is the chat
                                                         │   ├─ Store: SQLite (credentials, session ids, markers)
                                                         │   └─ Static PWA
                                                         └─ tailscale serve (TLS on the MagicDNS name)

Requirements

• A Mac that stays on (the service runs under launchd), with uv and tmux
• A Tailscale account, with the Mac and iPhone on the same tailnet
• API keys: Deepgram (streaming STT), Cartesia (streaming TTS)
• At least one coding agent CLI on the Mac, logged in: Claude Code and/or Codex CLI. Either alone is a complete install; every session is your existing CLI setup — skills, hooks, MCP servers and all
• Optional: a role-root skill on your global skill path launches every workstream into a role instead of a bare session — deploy-rw detects it and falls back cleanly if it's missing. ryan-scheinberg/harness is the reference implementation

Install

The install is one guided skill — run it from whichever CLI you prefer. It wires every engine you have (asking first), so installing through one still makes the other pickable from the phone.

From Claude Code:

/plugin marketplace add ryan-scheinberg/remote-workstreams
/plugin install remote-workstreams@remote-workstreams
/remote-workstreams:deploy-rw

From Codex:

codex plugin marketplace add ryan-scheinberg/remote-workstreams
codex plugin add remote-workstreams@remote-workstreams

then start codex and ask for $deploy-rw.

The deploy is run by the agent on your Mac. It confirms every system-touching action with you before running it, and it is safe to re-run — it doubles as repair. What it does:

1. Preflight: macOS, uv, tmux, and a durable git clone of this repo (defaults to ~/remote-workstreams)
2. Engines: detects Claude Code and Codex; with your OK, wires whichever second engine is present so both are pickable from the phone (the picker only shows wired engines), and points the planner/injector at the engine you installed with
3. Tailscale: detects it, guides install and login if missing, captures your MagicDNS name
4. Stores your two provider keys (Deepgram, Cartesia) in the macOS Keychain
5. Takes your 4-digit pairing PIN; only its scrypt hash is stored
6. Installs and starts the launchd service, verifies /healthz
7. Maps HTTPS on your MagicDNS name to the local service via tailscale serve
8. Prints a pairing QR code — open it on the iPhone, Add to Home Screen, enter the PIN, confirm Face ID
9. Runs an audio round-trip test (synthesized speech in → transcript → reply audio out) and reports the result

Security model

• Tailnet is the perimeter. The service binds to localhost; only tailscale serve exposes it, and only devices on your tailnet can reach it at all.
• Pairing, once per device: your 4-digit PIN + WebAuthn registration (Face ID); the passkey's public key is stored. Five wrong PINs lock pairing for 10 minutes.
• Login, every app open: one Face ID tap (WebAuthn assertion against the stored passkey) mints a session token held only in server memory (24h TTL) and only in a page variable on the phone — a Lock button, a reload, or a server restart ends it.
• Secrets live in the macOS Keychain (service remote-workstreams) — provider keys, and only a hash (scrypt) of the PIN. Nothing secret in config files.
• Passkeys are listed and revocable server-side. Lose a phone, revoke its credential.

Latency

Measured, not promised: every turn logs endpoint → transcript → first-sentence → first-audio timestamps. Replies come from a real interactive session, so expect the pace of a thoughtful colleague, not a kiosk — sentence-chunked streaming TTS starts speaking as soon as a reply lands, and barge-in (speak over the assistant and it stops) keeps you in control.

Development

uv sync            # install (Python 3.12/3.13)
uv run pytest      # test suite (no live API calls; SDK boundaries are mocked)
uvx ruff check .   # lint

Tailscale Funnel

tailscale funnel can expose the service to the public internet. This is documented as a pointer only and is unsupported: remote-workstreams's auth assumes the tailnet perimeter, and v1 has no public-internet hardening (rate limiting, lockout). Don't do it unless you understand exactly what you're removing.

License

GPLv3.