Everything Setu does, in plain words.

1. What Setu is

1. Setu is your own AI company. You describe the business, Setu proposes departments and AI employees, you approve who works, and every result comes to you for review before anything leaves the company.
2. Employees are people, not bots: each has a role, instructions, skills, its own memory, its own storage, and (if you allow it) its own computer and connected tools.
3. Nothing is sent, published, bought or changed outside Setu without your approval, unless you mark a tool as trusted.
4. What costs money: only live AI runs (with your own AI key or the platform key) and your plan. Practice runs, reviewing, editing knowledge and setting things up are free.
5. Help & guide at the bottom of the sidebar opens this guide at any time; the small help button on every page explains the page in front of you.

2. Start in five minutes

1. Voice: press the microphone beside “Tell your company what you need” (or beside the company description when you create a company) and speak in English or Hindi; your words are added to the box for you to check and send. Open any draft and press “Read this draft aloud” to hear it in the most natural voice your device has. Voice works in Chrome and Edge; nothing you say is sent to Setu until you press Send.
2. Set up by talking: when you create a company, press “Set up by talking”. Setu asks the company name, what it does and which starting point fits, listens to each answer in English or Hindi and fills the form; say it again if it did not understand. Nothing is created until you check the details and press Create company.
3. Just looking? “Explore the demo company” on the landing page gives you your own copy of a running company: edit anything, it is saved. Choosing a plan lets its employees run.
4. Team → “Plan my team” (open by itself while you have no employees) → describe the business in plain words. Setu proposes departments and draft employees; press “Activate all drafts” to put them to work. Drafts cost nothing until activated.
5. Settings → AI engine: choose practice mode (free, structured briefs only), or OpenAI, Anthropic, OpenRouter or any OpenAI-compatible API and paste your own key. The key is encrypted and never shown again. The Getting started card on Home opens this tab for you.
6. Settings → Plan & usage → set a company spending limit (the “Authorized limit”). Live AI work stops at that ceiling; practice mode needs no limit.
7. Settings → General → Interface language switches Setu's own menus and headings to Hindi on this device; “Language your employees write in” is a separate, company-wide choice for drafts and replies.
8. Home → type what you need in the box (“write a WhatsApp reply for a delivery question”). Setu suggests an employee, or you pick one; the result appears under Work → Tasks and, when ready, under Work → Approvals.
9. Example: type “Draft a 3-line WhatsApp reply to a customer asking where their order is.” Expected output: one task assigned to your support employee, and a few minutes later a draft reply waiting under Work → Approvals for you to accept or send back.
10. Limits of a chat request: each request carries a fixed allowance of up to ₹500 of the employee's own limit, and that allowance is a real ceiling. If it stops with “Needs a higher limit”, open the task and press “Raise this task's limit from … to … and run” (you confirm the new amount), or raise the company limit under Settings → Plan & usage. Limits are never raised on their own.
11. If nothing happens after a request: check that at least one employee is active (Team), that a key is saved (Settings → AI engine) and that the company is not paused (Settings → General → Operating boundaries).

3. Give your company its knowledge

1. Project memory: when you add knowledge or a document, choose “One project (objective)” and pick the project. Only tasks under that objective use it, so one client's brief never reaches another client's work. Everyone who can see company knowledge sees it, labelled with the project. Facts an employee learns from it stay with that project and are not added to the employee's memory for other work.
2. Knowledge & tools → Knowledge → “Write a procedure, a handbook, or custom knowledge”. A handbook is company-wide standing knowledge: vision, mission, policies, brand voice, how you work. Write it first.
3. A procedure is a step-by-step way of doing a job, for the whole company, one department, or one employee. Employees carry handbooks and procedures on every task and follow them.
4. Custom knowledge (price lists, product details, client facts) is used when it is relevant to a task. Tick “always include” if it must be there every time.
5. Upload files the same way: PDF or Word up to 2 MB, or text, Markdown, CSV and JSON. Scanned pages are read by text recognition, so check numbers afterwards. Employees never see anything outside their access scope.
6. Example: add a price list as custom knowledge, then ask “Quote 50 units of product A with delivery to Pune.” Expected output: a draft quote using your exact prices, with anything it had to assume marked for you.
7. Costs: writing and uploading knowledge is free; knowledge only costs anything when an employee reads it during a live AI run. Without knowledge, employees draft with assumptions and mark them for you.
8. Facts your employees learned from work wait here and on each profile for your approval (approve, correct or forget). When two sources or facts disagree, “Information that needs your decision” opens on this page and both records are kept out of work until you decide; it stays folded while nothing is open.

4. Shape each employee

1. Hire from a ready role: when you hire, “Start from a ready role” fills the job title, instructions and skills for 18 common roles (chief of staff, customer support, social media, SEO, content, performance marketing, email marketing, sales development, bookkeeper, HR recruiter, operations, research, data, product, software, catalogue, contract reviewer, executive assistant). Every one drafts for your review; change anything before saving.
2. Rearrange the team on Team: drag an employee onto a colleague to make them report to that person, onto “You” to report to you, or onto a department chip that appears while dragging. The employee stays active; the move is recorded in History. A move waits while the employee is in the middle of a task. The employee's Edit role & access form does the same with the keyboard.
3. Team → click an employee → “Edit role & access”. Instructions say what the job is; Skills say how to do parts of it (one per paragraph, “Name: how to do it”). The operating limit is the most one run may spend.
4. Import skills: on the profile, “Import skills from a file or GitHub”. Paste a GitHub repository, folder or SKILL.md address, or upload a file, read the preview, tick what to install. Installed skills become standing instructions you can edit later.
5. Memory: facts the employee learned from work wait on its profile for your approval. Correct the wording, approve, or forget. Nothing unapproved is used.
6. Storage: notes the employee keeps for later tasks, plus its computer files if it has a computer. You can edit, remove or add text files there and download other files. A computer with the desktop template also keeps the last three screenshots the employee took under Computer files, so you can see what it saw.
7. Access & tools: tick only what the role needs. Anything not in the list is not connected yet; connect it under Knowledge & tools → Connections first.
8. Outside agent (employee profile → Access & tools, owner only): hire an agent you already run elsewhere, such as LangGraph, CrewAI, n8n or a partner's agent, as this employee. Enter its OpenAI-compatible chat completions address, the model name it expects, its price per million tokens and its key (sealed and never shown again). Its tasks go to that agent; the budget, your review and the history stay in Setu, and the company's own AI key is never sent to it. Disconnect returns the employee to the company engine.
9. Standing duties (employee profile → Duties): give an employee a routine, such as a morning inbox check every day at 08:30, every Monday at 09:00 in your time zone, or every 4 hours. Each duty has its own spending limit per run, and a task is created and started on schedule; the result comes to you for review. While the last result is still open, that run is skipped; a paused company or employee also skips it.
10. Only on working days: tick it on a standing duty and it skips days off, company holidays and times outside working hours set under Settings → General → Work calendar (working days, holidays such as 2026-10-02, and optional hours, in the company time zone). The duty shows “Day off: …” and runs at its next time; duties without the tick run every day.
11. Example duty: “Every day at 08:30: read new customer messages, draft replies, list anything the owner must decide.” Expected output: one task each morning under Work → Tasks marked as a standing duty, with a draft under Work → Approvals. If a duty cannot start, you are notified (Settings → Notifications) and the reason shows on the duty.
12. Performance (employee profile → Performance): accepted first time vs after changes, acceptance rate, cost per accepted result, typical time to your acceptance, sends that left the company, customer messages handled, quality-review criteria met, and accepted results per week. Choose the period at the top.
13. Customer messages reach your employees: on the profile → Access & tools → “Connect incoming requests”, choose WhatsApp (Meta Cloud API) or Email (Brevo inbound parsing) and follow the steps shown. Each message becomes a task with the sender's address and earlier messages, with its own spending limit per message; the reply is drafted for your review and goes out only through an approved send. Work → Tasks shows an Inbox of these messages.
14. If an employee's task stops with “Needs access”, open Team → the employee → Edit role & access and tick the tool, or grant the employee on the connection under Knowledge & tools → Connections; then press “Grant access, then run” on the task.

5. Connect the outside world

1. Knowledge & tools → Connections → pick a card → Connect. Every card walks you through the steps: the account, the key with a “where to find it” guide, an automatic check, and who may use it. Keys are encrypted; connecting is free.
2. Cards available today: apps by sign-in (through Composio), Shopify, GitHub, HubSpot, email (Brevo), WhatsApp Business, a Google Drive folder, a computer for each employee, a read-only database, web search (Brave), Razorpay (reads only), a code runner, a web browser, an MCP server and Meta Ads. Cards marked “Not available yet” (for example Google Ads) cannot be connected yet.
3. “Connect an app by signing in” searches Composio's app catalogue: sign in on the app's own page or fill in its custom setup, then tick the tools. Reads run on their own; tools that change data need a separately reviewed action, and each change still waits for your approval.
4. Calendars, Gmail, Sheets and Google Ads come through “Connect an app by signing in”: search Google Calendar (28 tools: read events and calendars; creating or changing an event becomes an approved action), Gmail (read mail, create drafts), Google Sheets, or Google Ads (campaign lookup and customer lists only; performance reports need your own Google Ads developer access or a Google Ads MCP server added under Connections).
5. “Give employees a computer”: an isolated Linux machine per employee with files that persist between tasks (text files up to 200,000 characters, other files up to 2 MB, at most 60 files). No internet inside, no access to your machines. The machine needs no key when the platform key is set.
6. Computer templates: keep “base” (Linux with Python 3 and Node.js) for command-line work, or type “desktop” for a graphical desktop the employee can look at and use: it can take screenshots, click, type, press keys, scroll and open files from its working folder. The last three screenshots are kept on the employee's profile under Computer files. A base computer has no screen and says so when asked.
7. Example: with a desktop computer, ask “Open the sales spreadsheet in the work folder and tell me the total for March.” Expected output: a draft answer with the figure, and the screenshots it took visible on the profile. Web addresses are refused: there is no internet inside.
8. Shopify, HubSpot, GitHub, Brevo email and WhatsApp use a key from that service; the guide tells you exactly where to copy it. A read-only database connection runs one query at a time with a ten-second limit and at most 200 rows; nothing can be changed through it.
9. Software teams on GitHub (connect GitHub through “Connect an app by signing in”, then add ready actions): “Create a work branch” (for example fix/login from main), “Save a file to a work branch” and “Open a pull request”. You approve each exact request; Setu checks the branch right before sending, never writes to the main branch, and reads each result back from GitHub to confirm it.
10. “Everyone, including employees added later” gives the connection to the whole team and to future hires. Each granted employee receives a skill note explaining the tools.
11. Trusted tools: on a connected tool press “Trust: run without asking” and employees can use it without waiting for your approval. Leave it off for anything that sends, posts or spends.
12. Pause one connection: Knowledge & tools → Connections → Pause next to an account stops every read and action through it at once, including approved actions not yet sent, without removing it. Resume puts it back exactly as it was. Revoke removes it for good.
13. Webhooks (Connections page, bottom): a website form, Zapier, your shop or any script can hand a task to a chosen employee under a chosen objective. The token is shown once; copy it then. Each delivery becomes a task reviewed like any other work.
14. If a connection stops working (“Needs access” on a task): open the connection under Connections, run its check again or paste a fresh key, then press “Grant access, then run” on the task. You can revoke any connection there at any time.

6. Work, review, approve

1. Home shows what needs you first: “Needs your attention” groups stopped tasks by cause with the one action that fixes it. “Needs a higher limit → Raise this task's limit (confirm the new amount), then run”, “Needs access → Grant access, then run”, “Failed → Open the task and retry”, “Needs changes → Send back with notes”.
2. Company progress (bottom of the sidebar): how much of the company's real work is finished, how much waits for you and how much is stuck. Press it to open Work.
3. Work → Tasks lists every task with its state and the reason if it stopped. “Run with connected AI” starts it (costs AI usage); “Run local fixture” makes a free practice brief. A task's own limit is a ceiling the product never raises by itself: when the estimate is above it the task stops with “Needs a higher limit” and the dialog offers “Raise this task's limit from … to … and run”, which you confirm. A task assigned without a limit takes the employee's per-run limit at its first paid run (practice runs leave it unchanged) (recorded in History as a System change). A run above the employee's per-run limit stops until you raise that limit under Team → employee → Edit role & access. If the AI engine fails or does not answer, the task shows “Outcome needs checking”: the amount reserved for that attempt is held, and running it again counts the full reserved amount as estimated spending because the provider's final usage is unknown. Cancel task stops later model calls. Any model call or knowledge lookup already sent can still incur a cost; known usage is counted as measured, and an unresolved lookup keeps its estimate. An answer that arrives after cancellation is withheld.
4. Choosing the AI model: Settings → AI engine sets the company's model. When assigning work you can pick another approved model of the same provider for that task (for example a faster, cheaper one for simple work); its own approved price applies. A backup model (Settings → AI engine) takes over new runs automatically while the main model's last three attempts have failed, and the run history says so.
5. How the work was done: open any task and expand “How the work was done” to see each run (practice or connected AI, when, its state and measured cost). “View steps” shows what the employee looked up, what came back and what each step cost. People placed in a department see their department's runs.
6. If a run fails: open the task, read the reason, and press “Run with connected AI” again to retry. If it fails again, check Settings → AI engine (key and model) and the employee's access. After three attempts in a row end without an answer from the AI engine, new paid runs pause for ten minutes (the message says until when) so no more of your limit is held against an engine that is down; practice runs and scheduled duties carry on, and duties simply try again later.
7. Work → Approvals shows each draft exactly as it will be accepted. Accept it, or Request changes with notes; the employee revises with your notes next time. “Changes to connected accounts” below lists sends and account changes waiting for your approval, one by one.
8. Department sign-off (Settings → General → “Department heads sign off first”): a person you placed in a department as manager or reviewer signs off that department's drafts before anyone accepts them, and is emailed about them. The approval card shows “Waiting for sign-off from …” and then “Signed off by …”. Heads can sign off or request changes but not accept. An owner or administrator can still accept without the sign-off by writing why; the reason is kept and shown in Explain this decision.
9. Managers can review their team's drafts first (Settings → General; one to three rounds). The manager's verdict appears on your approval card. Managers cannot accept drafts or approve sends for you.
10. Explain a decision: open any accepted or sent-back draft (task → Output history) and press “Explain this decision”. It shows only what was recorded then: who decided and when, the reason (optional when accepting: write it in the notes box), the manager's note, the other versions and why they were sent back, the brief, limit and measured cost, and the sources and remembered facts exactly as they were used. Anything not captured says “Not recorded”. It lists what has changed since (a source revised or removed, the employee paused, the limit changed) with dates. “Reconsider this decision” starts a new task for the same employee with that record and your note; the earlier decision stays as it was and nothing is sent without review.
11. Work → Results keeps every accepted output with its version. Open it, download it, or press “Save as knowledge” so employees can use it in later work.
12. Work → History is the company audit trail: every action, approval, tool call and spend, latest 100 events, with details on each.
13. Run several companies (agencies, groups): with more than one company, All companies appears at the top of the sidebar and shows each company's drafts, sends and stopped runs, work in progress, accepted results, employees, customer messages, spend and plan, sorted by what needs you. Open a company from there to act.
14. Ready-made companies (Team → “Plan my team”, while you have no employees yet): D2C brand, Digital agency, Clinic, Real estate, Coaching and courses, and Software company (the large one: a chief of staff, four department heads and their teams, 11 employees). One click creates the departments, draft employees with skills, a handbook to fill in, procedures, objectives and standing duties; nothing is activated or spent until you say so.
15. Plans, add-ons, invoices (Settings → Plan & usage): your plan caps active employees and runs per period; add-ons add runs to the current period. Every payment produces a numbered receipt (a GST tax invoice once the seller's GSTIN is recorded) you can open and print; save your legal name and GSTIN under Billing details. Ask for a refund from the invoice list within 7 days of a payment, if fewer than 10 runs of that period were used.
16. Notifications (Settings → Notifications): Setu emails you, and can WhatsApp you through your own WhatsApp Business connection, when a draft waits for review, a send waits for approval, a run stops, a limit is reached, a customer writes in or a duty cannot start. Several things at once arrive as one message. The email carries a one-use link: open it on your phone, read the draft, and accept it or send it back with notes without signing in. If your company has not connected its own WhatsApp Business account, WhatsApp alerts come as a short message from Setu's own number (when Setu has one set up).
17. People (Settings → People): invite members with a one-use invitation link and see who belongs to the company.
18. Discussion on any task (open the task): comment as a person; mention an employee with @ and their name to ask them something. The employee's reply is a draft you review; once accepted it is posted in the thread in their name.
19. Employees hand work to each other: a draft can include work for a colleague. Once you accept the draft, a direct report's task runs on its own; a peer's task waits for you to start it. Each shows “Handed over by”.
20. Anything that changes data outside Setu (send an email, create a product, post a message) is a proposal you approve one by one, unless you trust that tool.

7. Objectives, workflows and schedules

1. Work → Objectives → “Plan an objective”: describe the outcome and what a good result looks like; a manager proposes the steps, who does what, and how much each may spend. Planning with a connected AI model uses your key and limits; the practice template is free.
2. Example: objective “Launch the Diwali offer”. Expected output: a plan with steps such as write the offer, design the post, draft the customer email, each with an employee and a limit, for you to review and save as a workflow.
3. Work → Workflows: publish a workflow before running it (editing creates a new version). Steps run in order and in parallel where possible; each step still comes to you for review. Dependent steps wait for accepted inputs; a decision step runs only the route you accept.
4. Schedules: put a workflow on a repeat, every day at a local time or at a regular interval, in your company's time zone. A previous run must finish before another starts; choose what happens to runs missed while the service was offline.
5. If a workflow step stops for a low limit, open its task and press “Raise this task's limit from … to … and run” (you confirm the amount); the rest of the workflow continues from there.

8. Money and safety

1. Every AI request reserves an estimated cost before it runs and settles the measured cost after. Limits exist per company, per employee (per run), per department and per task; nothing runs past them. Limits are ceilings, not prepaid money.
2. Settings → Plan & usage shows the authorized limit, what was used, what is reserved and what is available, plus every usage record by employee and task. Work → History shows every action, approval and tool call.
3. When the company limit is reached: raise it under Settings → Plan & usage → “Adjust limit”, then re-run the stopped task. When only one task is short, use its “Raise this task's limit from … to … and run” button instead.
4. The free trial lasts 7 days and needs no card; afterwards choose a plan under Settings → Plan & usage to keep live AI work running. Records, knowledge and practice mode stay either way.
5. Keys are encrypted on the server and never shown to employees. Revoke any connection from Knowledge & tools → Connections at any time; pause the whole company from Settings → General → Operating boundaries.
6. Settings → Advanced: export your organization as a portable file (private instructions, sources and keys are left out), import one, see the “Behind the scenes” health panel, and archive or delete the company.

What each page does

Try it

Explore the demo company Start free