From sign-up to a team that runs a function without you, in six steps. Every new organisation gets 200 free turns, so your first team can start straight away; after that, buy turns as you go or pick a plan.
Jugg.ai runs on your model provider account — we hold no model key of our own, so nothing works until you add one. Open Settings → Model provider, pick a provider and paste an API key. One key runs the whole team.
Teams work in three tiers: Genius directs, Smart manages, Fast does the hands-on work. In Settings → Model provider each tier has two dials. Thinking runs from Low to Max (whatever the chosen model supports): deeper thinking is more thorough and your provider bills more per turn, and because it takes longer, a turn at Extra high counts as 2 against your plan and at Max as 4. Context (Standard, Deep or Max) sets how much of a job's history agents keep in view — more history is re-sent, and billed, every turn. The defaults are set to get work done: Genius thinks hard with Deep context; Smart and Fast run at Medium with Standard context.
On the dashboard, Connections lists the services your teams can use — close to a hundred, grouped by kind: GitHub, Google (Gmail, Drive, Calendar), Microsoft 365, Slack, Notion, Linear, Jira, Stripe, Attio and Close CRMs, analytics, hosting and more, most connected with one click and a sign-in. Anything else: any remote MCP server, or any REST API you describe with an OpenAPI spec. Connections are shared with every team in your organisation by default.
How often a team asks you is up to you. Each team’s Ask me before setting is one of: any change; critical actions only (the default — reads and ordinary edits go ahead, while sending mail or messages, inviting people, deleting, and changing who can see something ask first); or nothing — fully autonomous. Reads never ask. Each connection can also set its own level — say, full access for your analytics, and ask-first for payments. Budgets, the activity log and the kill switch apply at every level.
Teams search the web and read public pages on their own, with your model provider’s own search (Claude, OpenAI and Grok each have one; DeepSeek doesn’t), at about a cent a search on your key. Searching only reads, so it never asks. On a fully autonomous team, anything that sends, deletes or shares is first checked against what the job is for, and waits for you if it doesn’t fit.
Docs is your written library — policies, product facts, runbooks — and every team reads it before it acts. Click Import to bring in what you have: your help centre or docs site (enter its address; start at /help to take only those pages), or upload PDF, Word, Markdown, text or HTML files. Importing again refreshes what changed and never overwrites a document someone wrote or edited.
Teams can draft new documents and corrections for you to publish, and ask you to import a site they need — both wait under Needs you.
Click New team and describe the function you want handed over, e.g. “Maintain our GitHub repo: triage new issues, fix small bugs, keep dependencies current.” The Workforce Architect proposes a team; adjust it and create it.
On the team page, the Mission is the standing goal the team works toward. Keep it concrete — what good looks like, what is off limits.
Give the team a job on the team page — a bounded first job is best (“read the README and propose three issues; ask me before filing”). The director breaks it into tasks, managers assign them, workers do them.
When a worker reaches an action its team’s setting says to ask about, that step pauses and shows up under Needs you on the dashboard. Approve or reject it there; the team picks up exactly where it stopped.
Autopilot mode on the team page decides who starts work:
A trigger puts the team on call. Give it “What to watch for” (“unread emails in the support inbox that need a reply”) and “What to do when there's new work”. A cheap Fast-tier check runs on a schedule and starts a job only when there is genuinely something new.
Meetings. Under Meetings, pick the teams, say what the table should decide, and watch their directors talk it through. You can add a note mid-meeting and download the transcript. The meeting ends with a summary, decisions and a job for each team that agreed one — started straight away unless you turn that off (a team set to ask before any change still asks). Directors can also call a meeting themselves.
Jobs in order. When one job needs another’s result — a draft that needs the data first — the meeting can say so: the second job waits, then starts with the first job’s report in hand, and is cancelled if the first one fails.
Connect a bot. In Telegram, message @BotFather, send /newbot and copy the token. An owner or admin pastes it into Settings → Telegram and picks the team new chats talk to. It’s your organisation’s own bot, so nothing else can message you as us.
Link your chat. Anyone in the organisation chooses “Link my Telegram” in Settings and sends the bot the one-time /link code it shows (it works once, for 15 minutes). Only private chats can be linked.
Use it. Type a job in plain words and the team starts on it, exactly as if you’d given it in the app — same budget, same approvals. /teams lists your teams, /team 2 switches, /status shows what’s running and what needs you, /unlink disconnects the chat. Owners and admins also get a message when a team needs a decision, finishes a job or stops; the decision itself is made in the app.
Connect. An owner or admin opens Settings → Slack and chooses “Set up Slack”. Copy the manifest it shows, create a Slack app from it, install it to your workspace, and paste its bot token and signing secret back. It’s your own app, and it only reads what’s said to it — mentions and direct messages, never whole channels.
Link and use it. Choose “Link my Slack” and DM the bot the one-time link code. Then mention it with a job in any channel it’s in. It replies in a thread, and the team’s questions and its result arrive there. Mention it in that thread to answer the question, or to steer the job while it runs.
Teams have people. On a team’s Settings tab, name its lead, approvers and members. Approvals and agents’ questions go to the lead and approvers; whoever gave a job hears when it finishes or fails. With nobody named, that’s the owner and admins. Mark a team private and only its people (and owners and admins) see its jobs, memory and documents.
Agents ask. When a decision is genuinely yours, an agent asks and waits — up to a day — for your answer, in the app, by email, in Telegram or in its Slack thread. Approving takes an optional note the agent reads, and you can fix the wording of an email before it goes, or tell the team to stop asking about an action you trust.
Steer. On a running job’s page, “Steer this job” sends the director guidance it acts on at its next step.
Work that arrives. A trigger can run at a time of day (“weekdays at 9am”), or start a job whenever its private URL is called or an email reaches the team’s own address. What arrives is treated as information, never as instructions.
Two separate bills. Your model provider bills you directly for tokens, at their prices, on your account. We bill a flat monthly plan for the orchestration — the memory, skills, approvals, budgets and recovery. A plan sets how many teams run at once (from 2 on Starter), how many agents can work at once across them (from 15 on Starter), and how many agent turns they get each month (from 5,800 on Starter).
What a turn is. One model call by one agent. A trigger checking for new work is a much smaller call and counts as a tenth of a turn; a turn at Extra high or Max thinking runs far longer and counts as 2 or 4. The allowance is a hard limit: when it is used up, your teams pause until the plan renews or you add a top-up (US$20 for 4,000 turns, paid up front; top-up turns never expire).
Pay as you go. No plan needed: buy a top-up and one team runs on it. Without a plan a top-up buys 2,000 turns rather than 4,000, so a plan is the better deal once you're using it every week. Prices are in US dollars, excluding tax.
Spend shown in the product is an estimate of your provider cost at list price. We never mark it up and never charge it; we show it so you can see where money goes and so the caps below can stop it. In our own runs on Claude, drafting a reply to a customer used 2–6 turns and about 5–8¢; a written report, 10–15 turns and about 10¢; big jobs split across many agents use more. Your provider's invoice is the source of truth.
Every model turn is checked against these limits before it runs:
The check reserves room for a worst-case turn, so a team stops at the limit rather than overshooting it. When a limit is reached the work pauses — nothing is lost. A team that hits its budget is listed as halted under Needs you on the dashboard, with a button to raise its budget and resume (the team page shows the same as an Out of budget card, with Raise budget & resume). When the month's turns run out, Needs you says so and links to adding a top-up. Nothing ever raises a budget or buys turns on its own.
Stuck? Email support@jugg.ai. See also what's new.