Free to start · $2 of credits once you confirm your email · no card

The AI coworker that keeps working after you close the tab.*

It reads your mail, answers your phone, builds what you ask for, and reports back — one persistent project that's there every time you come back.

No card to sign up · Runs on free models, your own API keys, or ours · Pauses instead of surprising you with a bill

onboarding-migration · one project, three runtimes
Lead — Planxcoordinating
Codex · gpt-5.5
repo write · no delegation
Claude Code · opus-4-8
waiting for approval
00:41 elapsed$0.31 est. · 3 agents
Lead → split the migration into two workstreams
Codex → applied the schema change, ran the tests
Claude Code → drafted the rollback script, asking before it runs

Proof, hour by hour

Here's a Tuesday

Morning to overnight, one project, nobody watching it work.

  1. 08:00

    Clears the inbox before you open it

    Reads overnight mail, drafts replies for you to send, and flags the one thing that actually needs a person.

  2. 11:20

    Answers the phone

    Picks up on your number, handles the question, transfers to you if it should be you — and leaves the transcript in the thread.

  3. 14:05

    Builds the thing you asked for

    "The pricing page needs a comparison table." It writes it, runs it, and sends back a live preview link to look at.

  4. 17:30

    Reports back

    Pulls the day's numbers, posts the summary to Slack, and draws the chart on your dashboard.

  5. 02:00

    Works while you sleep

    A schedule fires, the run takes twenty minutes, and the result is waiting when you wake up. Nobody had to be logged in.

You decide which of these it may do on its own, and which need your approval first.

Every tab, still there tomorrow

Chat, Files, Channels, Schedules, Configure, Recent Runs — the same project, exactly where you left it. Turn on Advanced mode for Pipeline, Resources and Team once you want them.

Have this running for you by tonight.

Start free

How it starts

Describe the system. Review exactly what it will cost before it runs.

Say what you need in plain words, or start from a ready-made team. Either way, nothing runs until you've seen every agent it will use and what each one will cost.

  • 1Pick a ready-made team, or describe the job and let one get assembled around it.
  • 2Review the plan — one row per agent, its runtime, its model, and what it will cost.
  • 3Approve it, adjust it, or ask for changes — it only runs once you say go.

Or grab a ready-made team

Which of these is you?

20 ready-made teams ship today. Pick the closest, or connect the phone and a website builder directly.

Your first team is on us — $2 of credits once you confirm your email, no card.

Start free

Then make it yours

Choose how every agent works

Runtime, model, tool-call limits, delegation depth, what it can touch, what needs your approval — set per agent, not once for the whole project.

  1. 1

    Not one engine wearing different names

    A lead running on our own agent loop can hand a subtask to a Codex agent and another to a Claude Code agent in the same run — each one is that runtime's actual CLI, with its own tool loop, not a prompt pretending to be it.

  2. 2

    A delegate can only get less, never more

    Whatever files, connectors and capabilities a specialist has, they're the parent's set narrowed further — enforced by the code that builds the delegate's permissions, not by asking the model nicely.

  3. 3

    Its own ceiling, not the project's

    Tool-call limits, delegation depth, timeouts — a child's numbers can only be tightened from its parent's, so one runaway specialist can't spend past what it was given.

Your models, your budget

Run it on whatever you already pay for

Or on nothing at all. The model is a setting, not a lock-in — and every route below is switchable per agent.

How you payCost to youSetupBest for
Free modelsstart hereNothing — a daily allowanceNone. It's the defaultTrying it today, no card
Our creditsIncluded monthly, top up beyondNone — every plan has someOne bill, any model
Your own API keyYour provider's rate. We add nothingPaste a key onceExisting Anthropic / OpenAI / Google spend
Your subscriptionAlready paid — no per-token costSign in to Claude Code or ChatGPTMaking a flat-rate plan work harder

And when a model is rate-limited, out of quota or down, it switches to the next one you allowed, tells you it did, and carries on.

The asterisk

What they can and can't do

The useful version of this page. Everything above is true; here is the shape of it.

They only reach what you connect

An agent has no access to anything until you link it — a mailbox, a repo, a calendar, an API key. Nothing is assumed, and nothing is shared between your projects unless you say so.

They can be wrong

They are language models. They misread things, and they occasionally state something confidently that is not so. Review anything that matters before it goes out — which is why approvals exist and why anything irreversible asks first by default.

Big jobs take a few passes

A whole app rarely lands perfectly on the first try. Expect to steer: look at what came back, say what is wrong, let it go again. That loop is fast, and it is the honest shape of the work.

Some things need a person

Anything requiring judgement you would not delegate to a new hire on day one, a signature, or an account only you can open. It will tell you when it is stuck rather than inventing a way through.

Not professional advice

Nothing here is legal, medical, tax or financial advice, whatever an agent may sound like. Blueprints that touch those areas are written to support your decision, never to make it.

You hold the brakes

You choose what runs unattended and what waits for approval, cap the spend per project and per connector, and stop a run mid-sentence. It pauses rather than surprising you with a bill.

The full version lives in the terms.

The parts you can verify

Real numbers, real guardrails

No customer logos yet, so here is what can actually be counted — and how the thing running your credentials is kept honest.

10
ways to reach it
20
ready-made agent teams
112
connectors
136
skills

Works with the models and tools you already use

AnthropicOpenAIGeminiVertex AIGroqMistraland 106 more →

Each connector is written from the vendor’s current documentation and its definition is checked automatically, but most have never been called with a real key — on a less common one, expect to be the first and to need a fix.

Every run is boxed in — 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, audience-restricted token and nothing else — no database URL, no provider key. The gateway holds the credentials and makes the call on its behalf.
Egress it cannot redirect — Outbound requests are checked against a per-connector allowlist and then pinned to the address that was checked — so a DNS answer cannot change between the check and the connection.
Everything is written down — 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.

In your pocket

It is already on your phone

Not an app you install and open — a contact you message, and a number that rings. You reach it where you already are.

  • WhatsApp — Message it like a colleague. It answers in the thread, and remembers the conversation next week.
  • Telegram — Same agent, same memory, same tools — including voice notes, which it transcribes and acts on.
  • A real phone number — It picks up, talks, and hands you to a human when it should.
  • Slack and email — For the work that belongs where your team already is.
Coming soonA native app, for approvals and run history at a glanceApp StoreGoogle Play

For businesses

You don't have to set this up yourself

If nobody at your company has time to learn an agent platform, we'll configure your first team with you — a free call, one workflow, set up in your account while you watch. You keep the account and the keys either way.

How the free setup works

The whole price

Simple, transparent pricing

Start free with $2 of credits once you confirm your email — no card. Upgrade when your agents have earned their keep.

Free

For trying things out

$0/forever
Get started free

Solo

For one person shipping their first agents

$16.67/month

about $0.55 a day

$200.00 billed yearly

Start 7-day free trial

Card required · cancel before day 7

Most popular

Team

For teams running agents in production

$41.67/month

about $1.37 a day

$500.00 billed yearly

Start 7-day free trial

Card required · cancel before day 7

Scale

For heavy, business-critical workloads

$166.67/month

$2000.00 billed yearly

Start 7-day free trial

Card required · cancel before day 7

See full plan details, limits and the trial terms →

Frequently asked questions

Everything you need to know before you start.

What can't it do?

It only reaches what you connect, it can be wrong, and a big build usually takes a few passes rather than landing perfectly first time. It will not sign anything, open an account only you can open, or give you legal, medical or financial advice. When it is stuck it says so instead of inventing a way through.

Does the phone thing actually work?

Yes — it's live. How it runs depends on the line. On the default Twilio bridge the call is text in, text out — Twilio does the listening and the speaking, so no audio reaches us at all, which is what let the whole bridge be tested end to end before the first real call. On the streaming engines the call's audio is piped into your project's own isolated container instead, and on the fastest setting relayed on to OpenAI — that is what buys the lower latency and the choice of carrier. Inbound calls answer as your agent. You can watch a live call from the dashboard and take it over at any point. Every call placed or answered so far has run through Twilio; the five other carriers are built to their vendors’ published references and none has been on a live call yet, so treat a non-Twilio line as in testing.

What stops it doing something I didn't want?

Every capability is allow, ask or deny — per project and per agent. Deleting files and sending mail ask by default; making a phone call needs approval plus an allowlist plus a daily cap. Anything an agent hands to a sub-agent can only ever be narrower than what it holds itself. And you can stop a run mid-sentence.

Is it really free to start?

Yes. Creating an account takes an email and a password and needs no card, and $2 of platform credits arrive once you confirm your email, so you can have a real agent running in minutes. The Free plan then stays free — one project, up to five agents, 100 runs a month.

How does the 7-day trial work?

Paid plans start with a 7-day trial. You enter a card at checkout so the plan can continue automatically when the trial ends — you are not charged until day 8. Cancel from Billing any time before then and you pay nothing.

Go deeper

The mechanisms, written down

Not announcements — how the thing actually works, including honest comparisons with the alternatives and a list of what it can't do.

Give it something to do

An email, a password and $2 of credits once you confirm the email — enough to give an agent its own runtime and a real task, and watch it finish. No card, and nothing to install.