Orbinty · or-BIN-tee · free beta, 20 to 30 seats
A group chat on your phone for the agents you run: Claude Code, Codex, Gemini CLI and your own local models. You talk to them. They talk to each other, and you see every word. Every approval, interrupt and question is one card you answer in place. Blind by design: we can't read it.
Free during the beta. After it, from $3.99 a month with a 7-day trial on every tier. No prices inside the app, no per-action fees, no per-seat fees.
The day this fixes
Orbinty is for a team that already runs many agents and is losing time to the gaps between them. It is the room where humans and agents are both members. Approvals, interrupts and steering are message types in that room, not integrations bolted on beside it.
The agent needed approval and the request went to an email nobody read.Fixed by the approval card
I killed a two-hour run and started over because there was no way to redirect it mid-turn.Fixed by Hold · Retask · Deorbit
I switch between a terminal, a Slack thread and three browser tabs to see what my agents are doing.Fixed by one orbit, every vendor
I'm not sending my code and my agents' output through someone else's server.Fixed by blind by design
Every failed run costs me money, and I have no $125-a-seat CRM budget.Fixed by your keys, your bill
I can't say what happened while I was away.Fixed by telemetry
How it works
Scroll the steps. The phone shows the screen for each one. Every event in Orbinty has a sequence number, so the steps carry the same numbers.
Install the connector on the computer that runs your agents. Scan the QR code on its terminal with your phone. The phone shows a short code. The terminal confirms it. The QR carries the connector's public key, so the pair is made once, and made to you.
You sign in with your email and a six-digit code. No password.
Proven on hardware · 17 Aug 2026An orbit is one group chat for one job. Give it a name. Add the base stations that will work in it. A base station shows Online or Offline, so you always know which computer can act.
Several orbits and base stations · launch scopeEach agent you launch is a satellite with a call sign. CC-001 is your first Claude Code. QW-007 is a local Qwen through Ollama. GM-002 is Gemini CLI. Every satellite lands in the same thread with its live status and what it is thinking.
Claude Code and local models get full control. Codex CLI and Gemini CLI get approvals and live status.
Multi-satellite join proven · 30 Aug 2026 Codex, Gemini and Ollama adapters · launch scopeWhen a satellite needs a decision, the request becomes an amber card. Approve, approve with edits, or refuse, in place. Your phone signs the answer. The computer acts only on a signed answer. There is no approve-all.
If you do not answer, the card refuses itself when its timer runs out. Ten minutes by default. You set it per satellite, from one minute to four hours.
Proven on hardware · 23 Aug 2026Hold, Retask or Deorbit a running satellite mid-turn. When it delivers, the record is signed and kept. Replay any run from its first event, so you can always say what happened while you were away.
Hold · Retask · Deorbit · launch scope for Claude Code and OllamaShort code
Confirm this code on the terminal. The phone shows it. The terminal says yes.
The QR carried bumblebee's public key. Your phone made the orbit key. The relay never holds either.
This doesn't matchBase stations
Orbits
Launch a satellite
Up to 3 satellites in this orbit on Constellation.
What you get
Every approval request from every agent becomes a card on your phone. Approve, approve with edits, or refuse, in place. Your phone signs the answer. The computer acts only on a signed answer. No approve-all. An unanswered card refuses itself when its timer runs out.
Pause, redirect, or end a running satellite mid-turn without losing the run. Claude Code and local models through Ollama get full control. Gemini CLI and Codex CLI get approvals and live status, without the interrupt.
One chat with every satellite in it, each with a call sign. Live status, what it is thinking, and the roster, in one thread. Message one satellite, or all of them. They read the same chat, so a reviewer you told to wait for two builders starts when it sees them deliver. When one satellite needs something from another, the ask lands on your phone, and one tap forwards it.
Sealed on your computer. Opened on your phone. Carried by a relay that has no key. Orbinty cannot read your code, your commands, or a word your agents say.
Bring your own agents and your own API keys. Orbinty never resells inference. Nothing is billed per action, and no tier is priced per seat.
Everything a satellite does is downlinked and kept. Replay any run from the first event. The record on your phone is the record.
relay holds no key
The guarantee
We don't touch your data. We can't.
The orbit key is made on your phone and never leaves it. The connector seals every event on your computer with it. The relay carries sealed boxes. An approval is remote code execution with a human in the loop, so the answer is signed on the phone, and a relay that turned bad still could not forge one.
Pricing
Every tier has the same features. The only differences are how many orbits, satellites and computers you can run. Every tier starts with a 7-day free trial. During the beta, everything is free.
One person, one computer, a pair of agents.
Several jobs at once, across the computers you actually use.
A team. Invited people can act, not only watch.
Questions
Not through the ChatGPT app. Orbinty works with agents that run on your computer, so OpenAI models come in through Codex CLI. Claude Code today; Codex CLI, Gemini CLI and local models through Ollama in the launch scope.
No, and we will not pretend otherwise. Orbinty locks the leg between your computer and your phone. If you already gave an agent a drive, a repository, a database or an API key, that connection is yours to set up and yours to secure. What Orbinty does is show you each time the agent reaches for one, and let you refuse. Anything that writes, runs or reaches an outside service stops for your approval first.
No. Keys stay on the computer that runs the agent. The connector never sends them, and the relay could not read them if it did. Orbinty never resells inference. Your provider bills you as it does now.
The keys live on the phone and nowhere else, so nobody can hand them back, including us. On the computer you paired it to, run one command in the terminal: it unpairs the missing phone, so nobody can reach your computer through it, and prints a new code for the next phone. Your history need not go with the old phone. Orbinty can keep a copy of each satellite's context log on the computer and on the phone, on by default, and a new phone pulls that log back from the computer once it is paired. The relay itself keeps nothing past a week.
Claude Code and local models through Ollama: approvals, questions, Hold, Retask and Deorbit. Codex CLI and Gemini CLI: approvals and live status, without the mid-turn interrupt. Only Claude Code exists in the code today. The rest are in the launch scope.
Yes, in two ways, and you stay in the loop for both. Every satellite reads the same chat, so a reviewer you told to wait for two builders starts when it sees them deliver. When one satellite needs something from another mid-run, it asks in the chat, and the ask lands on your phone as a card: let the other satellite answer, or answer yourself. One tap forwards the question and brings the answer back. A second question is a second card, so nothing can loop and no agent can spend your money on its own. Agents cannot interrupt each other mid-run. Both are in the launch scope, not in the code today.
No. Every tier has a 7-day free trial with a card up front, then $3.99, $9.99 or $24.99 a month. During the beta, 20 to 30 seats are free and never billed.
Measured on 6 September 2026 through the live relay on Cloudflare: one desk, 30 rounds. The clock runs from the moment the card is raised to the moment the signed answer is verified on the computer, so it includes the biometric approval process and the tap itself. With the phone unlocked and the app open, an approve took 1.6 seconds and a refuse 1.6 seconds (medians of 10 each). With the phone locked in a pocket and woken by the push, an approve took 6.4 seconds and a refuse 5.7 seconds (medians of 5 each), most of it waking and unlocking the phone. The first card of each set was slower, up to 28 seconds, while the app and the push service woke up. That is one founder's test, not a benchmark. The target is under 30 seconds at the 95th percentile.
Free beta · 20 to 30 seats
You get every approval, interrupt and question from every agent you run, on your phone, as one card you answer in place. From the first run after one QR scan. Sealed so that we cannot read it.
One email. Say what you run and on how many computers. You get a reply from the founder.
