Polycarp Sign in Start your free month

Getting started with Polycarp

Everything you do to set up your agent, in the order you do it — and what to do when something looks wrong. Your agent in the admin answers questions from this guide.

  1. Your account and signing in
  2. What your agent reads
  3. Testing it before anyone else sees it
  4. Correcting what it knows
  5. Its name, voice and look
  6. Handing off to a person
  7. Answers from your team
  8. Procedures: letting it act
  9. Customers
  10. Your team
  11. Automations
  12. Switching it on and choosing a plan
  13. Adding it to your website
  14. Running it day to day
  15. Billing, usage and your plan
  16. Your data: export, erase, retention, closing
  17. When something looks wrong

1. Your account and signing in

You create your account at polycarp.yohanun.com/signup: choose a plan, give your website, your email and a password, then enter your card on the next page (on Starter nothing is charged for thirty days). Your agent is created the moment that is done and starts reading your site. We email you a link to confirm your address — until you click it the agent can be built and tested but not switched on. If we set the account up for you, your welcome email carries the first password. You sign in at polycarp.yohanun.com/admin/login; Forgot your password? there emails you a reset link.

Your account is your business. It holds one or more agents — usually one for each website or brand you run — and the people who may sign in.

Signed in, you land on Today: usage this month, leads and escalations waiting for you, and the last week's conversations. The left rail lists your agents; each has six tabs — Inbox, Knowledge, Procedures, Automations, Settings and Test bench. On a phone the rail slides in from the left. Account holds billing, retention, export and closing.

2. What your agent reads

Your agent answers only from what it has read, and from the systems you connect for procedures (section 8). It never answers from the open web or from general knowledge, and when its knowledge does not cover a question it says so and offers a person. So what it has read is the whole game, and the Knowledge tab is where you decide it.

Reading your website

When your agent is created it reads your site: every public page it can reach from your home page, up to 200 pages, skipping pages that are too thin to be useful. Privacy and terms pages are left out unless you tick Include privacy / terms pages. The Knowledge tab shows the crawl's progress while it runs and, when it is done, a report of every page — learned, skipped as thin, or unreachable and why.

Read the whole site runs that again. Pages it already knows are kept as they are unless you tick refresh, which re-reads them. Use it when you have added several pages at once.

Adding one page

Add a page takes one URL on your own site and reads it now. Only pages on your site's address can become your agent's knowledge — the tab refuses anything else and says why. The same box refreshes a page it already knows.

Documents

Upload a document takes a PDF or Word file (up to 20 MB): a price list, a brochure, a policy. Your agent cites the document and the page number when it answers from it. Add by link takes a Google Doc shared as anyone with the link, a Google Drive file, or a direct link to a PDF or Word file; the link is remembered, so refresh re-reads the current version later.

3. Testing it before anyone else sees it

A new agent starts paused: it answers you, not your visitors. The Test bench tab is the real chat window, exactly as visitors will see it, but talking only to you. Ask what your customers ask: opening hours, prices, what you do and don't do, how to book. Every answer shows the pages or documents it came from, so you can see why it said what it said. Test conversations never count towards your usage.

If it says it doesn't know something you have on your site, the page is probably not in its knowledge yet — check the crawl report, then add the page. If it answers something wrongly, the source it cites is wrong or out of date: fix the page on your site and refresh it, or remove it.

4. Correcting what it knows

In the Test bench, every answer carries two buttons under the sources it used: refresh re-reads that page from your site now, and remove takes it out of the agent's knowledge immediately. You can do the same from the list on the Knowledge tab. Removing a page takes effect at once; the agent never answers from it again. A page you removed can be added back any time.

Your agent never invents facts to fill a gap. If a question keeps coming up that your site does not answer, the fix is a page on your site (or a document) that answers it — then add it, or approve the answer your team already gave (section 7).

The Knowledge tab also shows where what your agent knows disagrees: two facts in your content that can't both be true, such as two different prices for the same thing. Each one names both sources, so you can fix the page that is wrong.

In the Inbox, any answer that cited sources can be opened to explain it: the passages it used, the page each came from, and whether each still stands or has been replaced since.

5. Its name, voice and look

On Settings:

6. Handing off to a person

When your agent cannot help, or a visitor asks for a person, it offers one. If the visitor agrees, the chat asks for their name, email and message and emails the whole conversation to your Escalation contact (on Settings) with a reference. It also appears under Leads & escalations on Today and in the agent's Inbox.

Open the conversation from the Inbox to read it and reply as the team: your reply goes to the visitor by email and appears in their chat window, and the agent steps back while you handle it. Hand back to the assistant when you are done, or it hands back by itself after thirty quiet minutes.

7. Answers from your team

When your agent couldn't answer a question and someone on your team did, the Knowledge tab offers that answer back under Answers from your team, drafted as a short entry for your knowledge base. Where nobody has answered yet, it asks you the question instead. Approve an entry as it is, edit it first, or dismiss it. Once approved, your agent answers from it and cites it as coming from your team; you can remove it again at any time. Look for new suggestions checks for new ones now.

Nothing a visitor says reaches your agent's knowledge on its own. Only what you approve does.

8. Procedures: letting it act

A procedure lets your agent do something, not only answer: look up an order, issue a refund. You set it up on the Procedures tab in three parts.

The model proposes each step; the rules are checked in code before anything happens. Every action has a limit. Up to it, the agent acts. Above it, or with a limit of zero, the request waits on the conversation for someone to Approve or Reject, and the answer appears in the visitor's chat. Every step is recorded on the conversation. Try it with our demo shop sets up a working example in one click.

Shopify

Create an app for your store in Shopify's Dev Dashboard, give it read_orders, write_orders and read_customers, release it and install it on your store. Then paste its client ID and client secret under Connect your Shopify store. Your agent gets two procedures: where is my order and refund an order, the refund starting at a limit of zero so every refund waits for you until you raise it. A signed-in customer can only ever reach their own orders; an anonymous visitor must give the email the order was placed with, and can't be refunded.

Actions during your free month

On Starter's free month, procedures run in full in the Test bench and on the demo shop, and lookups work for visitors. Actions for visitors stay off until an owner presses Turn on actions on the Procedures page, beside a list of each action's limit to check first. From then each action completed is €0.25; the free month still covers conversations.

9. Customers

Customers in the menu lists everyone your agents know: people your site signed in, and people who left their email for your team. Anonymous visitors stay anonymous. Once a conversation goes quiet it is read for what it shows, such as ready to buy, wants to book, unhappy or might cancel, with a one-line summary, and for what the customer said about themselves. Each customer's page has What we know, which your team can add to or correct, and their conversations. The agent uses what you know only when your site has signed the customer in. Today opens with the customers who need you, and an automation can tell Slack, email or your own system the moment a signal appears. An owner can erase a customer: every conversation, on every agent, and everything known about them.

10. Your team

The Team page lists everyone with a login to your account. An owner adds a person by email (they get a link to set their own password), changes their role, or removes them. Owners run the account: billing, the team, turning actions on. Members do the work: the inbox, replies, knowledge, procedures and approvals. Every reply shows who sent it, and every approval who gave it. Can approve sets how large an action each person may approve when one waits: any amount, up to a figure, or not at all. It is checked when they press Approve; above their limit the action keeps waiting for someone who can, and the refusal is recorded on the conversation. Anyone on the team can reject. Your plan includes 3, 5 or 10 people; each one beyond is €10 for any month they held a seat. Someone removed can no longer sign in at once, and their name stays on what they did.

11. Automations

On the Automations tab, choose a moment and where to tell. The moments: a visitor asks for a person; a visitor mentions a word you choose; the agent couldn't answer a question; a signed-in visitor starts a conversation; any new conversation starts; an action is waiting for your approval. Each can post to Slack, send an email, or call your own system with a signed webhook. An automation fires once per conversation (an approval alert, once per request), and Test sends a marked sample. Only you set automations up; nothing a visitor says can create or change one.

12. Switching it on and choosing a plan

Your plan and card were set at signup, so switching on is one switch: Active on Settings. On Starter your first thirty days are free and your card is first charged when they end — we email you a week before; cancel from your account page before then and you pay nothing. Growth and Business are billed from the day you signed up. Every plan continues, billed on the same day each month or year, until you cancel.

Full prices and what counts as a conversation are on the pricing page, and how they compare with other AI agents on our price comparison. Your account page shows the plan you are on and your usage as it happens.

13. Adding it to your website

One line, just before </body> on every page where you want the chat:

<script src="https://polycarp.yohanun.com/embed.js" data-bot="your-agent" defer></script>

Your welcome email has the exact line for your agent. Until the agent is switched on that line shows nothing to visitors, so it is safe to add early. The chat runs in its own frame: your site's styles cannot break it and it cannot read your page. If your site is built with a page builder, paste the line into its "custom code" or "footer scripts" setting.

Signed-in visitors (optional, off by default): if your visitors sign in on your site, the agent can recognise a signed-in visitor — remember their earlier conversations across visits and devices, and know the account facts you choose to send (their name, their plan). Turn it on under Signed-in visitors on Settings; it shows your developer the two lines it takes (a signed token on your server, and the token on the embed tag). A signed-in visitor's conversations and profile are walled to that one account. The facts you send become part of what the agent holds for that user, so send only what it should know and make sure your privacy notice covers it.

14. Running it day to day

15. Billing, usage and your plan

A conversation is one visitor thread your agent answers at least once in a calendar month, however long it runs. Your own test chats never count, and nothing a person on your team answers counts. The number on your invoice is the number you have been able to see on your account page all month.

An action a procedure completes (a refund issued, a change made in your system) is €0.25, on every plan. Answers, lookups and handoffs are never charged, nor is anything refused, rejected or failed, or anything in the Test bench. Each one is recorded on its conversation; the account page shows this month's count.

Beyond your plan's allowance, conversations continue at the published overage rate, invoiced on your next bill — unless you turn on the monthly stop on your account page, which pauses new conversations at the allowance (conversations already under way are never cut off). A daily ceiling derived from your plan stops a runaway; it never interrupts a conversation in progress.

Your account page holds the rest: Update card, the plan and interval (changes apply now, prorated on the next invoice), the monthly stop, Cancel, and every invoice with its PDF. A cancelled plan runs to the end of the period you have paid for; your account page shows the date. After that the agent comes off your site, nothing is deleted, and choosing a plan again brings it straight back. If a payment fails we email you; Stripe retries; after seven days the agent pauses until it clears, and nothing is deleted.

16. Your data: export, erase, retention, closing

Each agent's knowledge and conversations live in their own walled space on the platform, hosted in the EU. The privacy policy and data processing agreement say exactly what is held and who processes it.

17. When something looks wrong