Agent Sidecar

Repurposing Spotify's discontinued Car Thing into a physical human-in-the-loop control surface for autonomous coding agents.

AI coding agents like Claude Code can work autonomously for long stretches — but you still have to babysit the terminal to catch the moment they finish, fail, or need your permission to do something. Agent Sidecar moves that attention/approval loop onto a dedicated physical device on your desk.

You start a coding task. The Car Thing lights up showing the agent is working. When the agent hits something that needs a human — "run this database migration?" — the screen flips to an attention state. You approve or deny with a physical button. The answer travels back and the agent continues. No terminal-watching.

This is not a Stream Deck clone, a Spotify controller, or a tiny second monitor. The product is specifically a human-in-the-loop interface for autonomous agents.

  ┌─ NORMAL ──────────┐   ┌─ ATTENTION ───────┐   ┌─ DONE ────────────┐
  │ CLAUDE CODE        │   │ ⚠ ACTION REQUIRED  │   │ ✓ TASK COMPLETE    │
  │ Market Edge        │   │ Run DB migration   │   │ +184  -37          │
  │ ● Working  02:14   │   │ [1]APPROVE [2]DENY  │   │ Tests 24/24 ✓      │
  └────────────────────┘   └────────────────────┘   └────────────────────┘

Status

🟡 Early / research validated. Feasibility and architecture are done (docs/ARCHITECTURE.md); Phase 0 (the riskiest unknown) is in progress. No V1 yet — this repo is being built in the open, phase by phase.

The core insight

The whole project hinges on bidirectional approval being possible without hacks. It is, on both ends:

• Claude Code exposes an HTTP PreToolUse hook that POSTs a tool request to a URL and waits for the response to decide allow/deny. That response channel is how a physical button press becomes Claude's answer — no terminal scraping, no synthetic keystrokes. (hooks reference)
• The Car Thing, via DeskThing, gives a web (React) runtime on the device, a server↔device transport over USB with reconnect, and a configurable knob/button mapping — so we write an app, not firmware.

 claude (your terminal)
   │  PreToolUse hook (POST, waits for reply)
   ▼
 Sidecar daemon ──WS──►  DeskThing app on Car Thing
   ▲   holds the hook request open           │ ⚠ shows the request
   └──────────── button press ◄──────────────┘ you approve / deny
 → daemon answers the still-open hook → Claude proceeds

See docs/ARCHITECTURE.md for the full feasibility report, component design, event protocol, phases, risks, and V2 plans.

Phase 0 spike — the make-or-break measurement

A human takes seconds-to-minutes to approve. Everything depends on how long a Claude Code hook will hold its response open before giving up. spike/phase0-hook-timeout/ measures exactly that against the real CLI.

cd spike/phase0-hook-timeout
./run.sh 5000       # hold 5s, confirm the mechanism
./run.sh 120000     # push toward the ceiling

The daemon-side round trip (hook → server holds → returns permissionDecision) is already verified. See the spike's README for current results.

Planned architecture

TypeScript monorepo (npm workspaces):

Path	Role
packages/protocol	Shared event/action schema (the normalized agent protocol)
packages/agent-core	Framework-free state machine, approval registry, audit log
apps/daemon	Local HTTP (hooks) + WebSocket (device) server — the brain
apps/device	DeskThing app: the on-device React UI
adapters/claude-code	Hook config + payload→event normalization + risk rules
tools/fake-agent	Scripted event emitter — develop with no real agent
tools/virtual-carthing	Run the device UI in a desktop browser — develop with no hardware

agent-core and protocol stay dependency-free so "add Codex" or "add a phone client" is a new adapter, not a rewrite.

License

MIT © 2026 Aaron Barke