How Choreo works: from a sentence to an always-on agent team
Choreo is an AI agent platform built around one idea: the project, not the chat, is the persistent thing. You describe an outcome once; a team of agents is provisioned around it; and that team keeps working — on schedules, over your messaging channels, on a real phone line — whether or not you have a tab open. Here is the whole pipeline, start to finish.
You describe the outcome; Choreo proposes the team
Setup is a conversation, not a form. Say what you want handled — "triage my support inbox," "research competitors weekly," "build and run my site" — and the assistant proposes a concrete plan: which agents, which models, which connections, what it will cost to run. Nothing executes until you approve the plan, and every choice on it (model, billing source, connector) is yours to change before you do. If you'd rather not start from scratch, 20 ready-made teams — support desk, research team, recruiter, chief of staff — install in one click as a starting point you can reshape however you like. And if nobody at your company has time to do the describing, we'll do this part with you on a call, free.
Each agent gets a runtime, a model, and a sandbox
This is the unusual part. Most platforms run every agent on one in-house engine. Choreo gives each agent its own runtime: our own loop, OpenAI's Codex run as-is, and Anthropic's Claude Code run as-is — real binaries, not imitations, and the right one changes per agent, not per project. Bring your own API keys at no markup, use platform credits, or run agents on the Claude Code or ChatGPT subscription you already pay for.
Whatever the runtime, the containment is the same: every run is sandboxed with gVisor, a userspace kernel between your code and the host. Containers never hold your keys — a worker gets a short-lived token and the gateway makes credentialed calls on its behalf. And every capability an agent has — files, code, shell, email, phone, external APIs — is set to allow, ask, or deny, per agent. A sub-agent's permissions can only ever be narrower than its parent's.
It runs when you're not there
An installed team is reachable and schedulable from day one:
- Schedules. Cron-driven runs with four delivery modes — announce the result on a channel, store it in the dashboard, POST it to a webhook, or run silently. Anything a scheduled run would need to ask permission for is denied unless you pre-approved it; some things can never be pre-approved at all.
- Ten channels. WhatsApp, Telegram, Slack, Discord, email and more — message your agent the way you message people. Reads your email.
- A real phone number. Answers your phone - live, in production. You can watch a live call and take it over.
- Memory. Durable facts stored deliberately and recalled by relevance — the recall formula is published, and everything stored is visible and editable in your account.
- Research. Searches the web and reads what it finds, with results fenced as untrusted data rather than trusted instructions.
It reports back — and everything is on the record
Overnight work announces itself where you asked: a summary on WhatsApp, a reply in Slack, a result in the dashboard. Behind that, every model call, tool call and connector request is recorded append-only, with what it cost — so "what did it do at 3am" has an answer you can read, not a shrug. The honest close, because we publish limits next to features: agents only reach what you connect, they can be wrong, and irreversible actions ask by default. The full containment story is on the trust page, and what always-on actually means has its own write-up. The fastest way to understand it, though, is the product itself: describe one job you'd hand off — free, $2 of credits once you confirm your email, no card — and watch the plan it proposes.
See it for yourself — free, $2 of credits once you confirm your email, no card.
Start free