This is the demo: made-up data, nothing is saved. Join the beta to connect your own agent. Join the beta
Settings
Connect your agent
Four steps, once. Then Rokan checks new conversations in batches (a few times a day at most), tells you when some went wrong, and brings you the tested fix to approve.
51 conversations received (last: Nov 20, 13:54 UTC), 0 not checked yet. Last check: Nov 20, 14:40 UTC.
✓ What your agent is told
Linked to acme/front-desk. Rokan read the prompt from agent/desk.py:12.
Replace the prompt by hand
✓ Send your conversations
Create your address and send every finished conversation to it. In Vapi, paste it as
the Server URL. In Retell, as the webhook URL. In ElevenLabs, as a post-call webhook (transcription). From your own code, the pull request adds a
one-line sender. From anything else: curl -X POST ADDRESS --data-binary @run.json.
Add /customer-name to the end to say whose customer the conversations are from.
✓ Where Rokan tells you
PagerDuty: type pagerduty: and the Events API v2 integration key of the service to page (Services, Integrations, Events API v2).
✓ How much Rokan does by itself
The second choice covers one thing only: a rule someone on your team already approved, word for word, when the same problem shows up at another customer and the rule passes the test on that customer's past conversations too. You are told each time (it needs an alert channel) and can undo it with one click. Anything new always waits for a person, and after 40 rules Rokan goes back to asking.
Your own checks
Rules in your words that every conversation is checked against, on top of what Rokan finds by itself: one per line, for example "never promise a refund" or "always offer a callback when the calendar is full". Only the checker reads them; your agent is not changed.
Start from your past calls (optional)
Already running on Vapi, Retell, ElevenLabs, Langfuse or LangSmith? Paste a key once and Rokan reads your last calls now, so the first check runs on real history. The key is used for this one read and never kept. Read-only.
Connect your coding agent (optional)
One line in Claude Code (shown below); for Codex or Cursor, add the same URL and header in their MCP settings. Your coding agent finds where your agent's conversations live, asks you before sending anything, sends them, and can test a prompt change on your past conversations before you ship it. What the token allows: read this team's checks, findings and fix tickets; send conversations; set the agent's prompt (the old one is kept and you are told); start a check or a test; ask a question. It can never approve, ship or delete: a person does that here.
Not connected yet.
This line contains a secret. It is shown once: copy it now and paste it into your terminal where your agent's code lives. Don't paste it in chat, a ticket or a commit.
One agent per customer (optional)
Running a different agent for each of your clients? Give each client its own prompt.
Conversations sent to your address with that client's name at the end (.../in/<token>/acme-dental)
are judged against that client's prompt; everyone else against the prompt above. Approved rules apply to all.
How an approved fix reaches your agent
Two ways, and you can use both. If your repo is linked, each approved rule
also arrives as a small pull request that adds it to the prompt in your code. And your agent can
read the separate rules address (shown with your ingest address; it can only read) when it starts
and add the text it returns to its instructions, so nobody touches the prompt. An agent
that serves one customer adds ?customer= and its name, and gets only the rules for everyone
plus that customer's own. A fix
that holds a link, an email or a customer fact always waits for a person. Rejected changes are
never proposed again.
Checks Rokan suggests for your agent
Rokan read the prompt your agent runs with and drafted rules every conversation should pass. Nothing is added until you pick it.
- Never state opening hours that differ from the business's own records
- Never promise a service the business does not list
- Offer to pass a message to the team when the answer is not in the records
- Hand off to a person when the caller asks for one
Or try it on a file (no connection needed)
1. Your agent's conversations
20 to 200 conversations in whatever you already export: Vapi, Retell, LangSmith, Langfuse, OpenTelemetry or plain JSON / JSONL. Up to about 9 MB.
2. What your agent is told
Paste the system prompt your agent runs with, including the facts it is given about each customer (hours, services, prices, rules). Rokan judges every answer against what you paste here, so a fact you leave out can look like something the agent made up.
The sample is 6 made-up conversations from 2 made-up businesses. It runs the same way your own conversations would, and counts as one of your checks for today.
Delete
Removes every conversation, report, rule, link and address of this team, now.
Sign out
Your access code keeps working. "Everywhere" also ends this team's sessions in every other browser.
From your own system
Post conversations from CI or a nightly job with the same access code:
curl -X POST https://deploy.rokan.computer/app/fix \
-H "Authorization: Bearer $ROKAN_CODE" -H "Content-Type: application/json" \
-d '{"runs": "...export text...", "name": "export.json", "prompt": "..."}'
# -> {"id": "..."} then: GET /app/r/<id>.jsonYour data
Names, phone numbers, emails, addresses and card numbers are removed the moment a file arrives, before anything is stored or analysed. The file itself is deleted once read. Analysis runs on Claude through Anthropic's API, which does not train on API data. Everything is deleted within 30 days, or now with the button above. If you connect your agent, the conversations it sends are kept the same way (scrubbed, 30 days at most) together with the prompt you saved, and the same button removes them. Nothing changes in your agent unless you approve a fix, or you have chosen to let Rokan apply again a rule your team already approved (off unless you turn it on).