CARAPACE
Done, and nobody knows.
An agent that finished forty minutes ago is forty minutes of shipping you lost. The model was never the slow part — you finding out was.
The AI agent control app
Stop babysitting your agents.
Your agents can work for an hour without you — then sit silent until you walk back to a desk. CARAPACE flips it: finished agents stand in line to talk to you. They report out loud wherever your phone is, and your spoken answer sends them straight back to work.
iPhone · chat, push-to-talk voice and vision free forever · Runtime Control (roster · queue · hands-free) $4.99/mo or $39.99/yr, 7‑day trial
Works with
Claude Code
Codex
OpenClaw
HermesVision
Point it at the world.
Ask anything.
Your agents live in terminals. Vision hands them your eyes. Open the camera and the thing you're looking at — the broken screen, the full fridge, the street you're standing on — becomes the thing you're talking about.
- The answer lands on the view. Ask out loud and the reply overlays what you're seeing, live — that's the recording on the right, mid-walk through a theme park.
- “What can I make from this?” Snap the fridge, ask, and cook what it says. The error on your other monitor works the same way — show it, don't describe it.
- SCAN — a 15‑second burst from a few angles for “what is all of this?” It lands in a lane as a real conversation you can keep going.
- Photos ride the chat like any other message and stay in that lane's history.
The bottleneck moved
The agents got fast. Noticing didn't.
Every serious agent setup hits the same wall — and it isn't the model.
Babysitting isn't supervision.
Cycling through terminal tabs to check whether anything needs you is a job. It's the phone's job now.
Chained to the chair.
If answering an agent takes a keyboard, your agents only move while you're at one. Dinner, the drive, the gym — all dead air. It doesn't have to be.
The mechanic
Runtime control, in three moves.
Three moves, and the gap between “agent finished” and “you knew” goes to zero.
01
Every agent becomes a lane.
Point CARAPACE at the machines you already own — the laptop on your desk, the mini in the closet, the server that never sleeps. Every agent on every one of them lands in a single roster, sorted into In line, Working and Resting.
02
Finished work gets in line.
A lane that needs you doesn't rot in a notification you'll never open. It takes a place in the queue — #1, #2, #3 — and reports out loud, one at a time, in the order they finished.
03
You answer. It's dispatched.
Say the reply out loud. It lands in that exact lane, in that exact conversation, and the agent picks the work straight back up. The next lane in line starts talking. The laptop never opened.
You already SSH in. Here’s what that can’t do.
A terminal shows you state when you go look. CARAPACE speaks the moment work lands, ranks four finished lanes so you take them one at a time, and carries a spoken answer back into that exact conversation — phone in pocket.
A lane is its own session.
Each lane is an agent session CARAPACE starts on your machine — its own conversation, its own history, running as you with the sign-in your agent already has. It is not an attach to the terminal window you left open: your terminal keeps its session, the phone gets its own, and neither steps on the other.
Tool approvals reach your phone.
When a lane hits a permission prompt — run this command, edit this file — the gateway raises it on your phone as a tappable decision (Claude Code and Codex today). Nothing is silently approved: a decision you never answer declines safely, and your agent’s own permission config remains the ceiling.
Why a phone, not another dashboard
Dashboards make you watch.
A remote lets you command.
Most agent tooling is a status page — one more screen you have to go look at. CARAPACE is built the other way around: the work comes to you.
- It speaks first. A dashboard shows state and waits to be read. CARAPACE announces the moment work lands — you never poll, you never refresh.
- It goes where you go. A browser tab is still a desk. The Lock Screen, the car, a walk with headphones — command posts you already occupy.
- It answers back. Reading status isn't control. Speak, and the reply is dispatched into that exact conversation — the loop closes without you touching a keyboard.
- It isn't loyal to a vendor. Claude Code, Codex, OpenClaw and Hermes in one roster — switch models per lane, mix runtimes per machine, keep your options open. That's the recording beside this: all four tabs, one line.
Parallel work
Parallel, without the tab chaos.
Dispatch four tasks across four lanes, then walk away. They come back ranked and spoken, in the order they finished — nothing waits on you noticing, so more of it ships.
- Four tasks, four lanes, one line. Same machine or four different ones — the roster doesn't care where they run.
- Switch a lane's model from the row. Change the model, rename it, quick-chat it or hide it, without opening anything.
- Real work goes through this thing. The recording beside this asked a lane for a benchmark chart. It drew the PNG in code and the finished image landed straight in the chat — files, photos and results ride the conversation.
- Setup is one line you paste into the agent itself. It reads its own instructions and installs its own leash. Developers tend to like that part.
- Same Wi-Fi, your own Tailscale, or your own HTTPS. Setup detects what your machine already has, names its choice, and tells you before it changes anything. The gateway answers only to a key that never leaves your machine.
- Mac, Linux, a VPS or a Raspberry Pi — headless is fine, and it comes back on its own after a reboot.
Four refactors on one box.
Split a migration across four lanes on the machine under your desk. They finish minutes apart and report one at a time, so you get four clean handoffs instead of four alerts at once.
Research reads while the tests run.
A research lane works through the docs on the mini at home while test and typecheck lanes grind. Whichever finishes first gets your ear; the others never stop.
The deploy lane speaks when it's green.
A watcher sits silent on the always-on machine through the whole build. The moment it goes green it takes #1 and tells you — in the car, on a walk, wherever the phone is.
Voice everywhere
Hands never leave the wheel.
Turn on AUTO and the phone runs the room. Every lane that finishes reports out loud, announced by name, one at a time — and the mic opens for your answer the second it stops talking.
- Live captions on the Lock Screen. Who's talking, what they just said, and how many lanes are still in line — on the Lock Screen and in the Dynamic Island.
- Replies duck your music instead of killing it. The track comes back on its own when the lane stops talking.
- SPEAK / SILENT is one tap when the room needs to be quiet. Replies queue up and wait for you.
- Bring your own voice. Apple's system voices work with no key at all — or use your own from ElevenLabs, OpenAI or xAI. The key stays in your phone's Keychain.
- As many machines as you like in the one roster, with favorites, nicknames and per-lane control.
Yours, actually
No middleman. Not even us.
CARAPACE hosts no AI, proxies no traffic and keeps no copy of your conversations. The phone talks straight to your machine — your Mac, your server, any box you own — over a network that's already yours.
- Your hardware, your API keys, your models. The agents are the ones you already run.
- If you can paste, you can set it up. About five minutes: copy one line, paste it into Claude Code, Codex, OpenClaw or Hermes, scan the code it prints. That's the install.
- The app is a remote. Your computer does the work; the phone just tells it what to do and listens to the answer.
- Nothing is reset or replaced. Timestamped backups, additive config merges, and your existing agent sign-ins, memory and history left alone.
# paste into Claude Code, Codex, OpenClaw or Hermes
› set this up: https://carapace.info/install
# it names its own runtime, fetches a self-verifying
# installer, and shows the QR only after it checks itself
INVENTORYPLANYOUR APPROVALSCAN THE CODE
Chat, push-to-talk voice, vision and pairing as many machines as you like are free, permanently. Runtime Control — every machine in one roster, the finished-work queue, and hands-free auto-announce — is CARAPACE Pro: $4.99/month or $39.99/year, 7 days free.
You already run the agents.
Take command.
Five minutes from pasting one line to your first spoken report. Chat, voice and vision free forever — Runtime Control when you're ready to run the room.