Make the docs answer the first question. Answers from your docs. Team replies in Slack or Telegram.

Where developer support breaks down

Developer questions often include details that ordinary help content misses: the SDK version, request method, response body, or order in which a secret was rotated. A quick answer that ignores versioning can make a small issue worse. A human engineer can resolve an unusual edge case, but should not have to repeatedly point users to the authentication page or explain the published retry policy.

Examples include “Where do I pass the idempotency key?”, “What does error code 429 mean?”, “How do I rotate my webhook secret?”, and “Does this endpoint accept a cursor or an offset?” Those questions are answerable only when the relevant reference is public, current, and unambiguous. A missing response example should result in a clear admission and a handoff, not invented code.

• Keep version numbers and deprecation dates next to the affected API behavior.
• Separate general error guidance from account-specific incident diagnosis.
• Use examples that match the SDK and language version you support.

Grounded answers, with an engineer in reserve

Connect public documentation such as Docusaurus, Mintlify, GitBook, Next.js docs, or another crawlable site. Include API reference pages, authentication, pagination, rate limits, webhook signatures, and known error codes. MivoChat answers using those pages; it does not connect to a private repository or execute a suggested request. Review the real questions developers ask against the pages you have made available.

When the docs cannot answer why one customer’s request fails, route that conversation to Slack or Telegram. The teammate sees the conversation and can ask for a sanitized request ID or a minimal reproduction. Reply in the Slack thread or Telegram topic and the visitor receives it in the widget. The teammate can take over, then hand the exchange back to AI or close it after the investigation.

Pay for replies, not a fictional one-question session

Starter costs $29/month for 1,000 AI replies, 3 seats, and one site. Pro costs $99/month for 10,000 replies, 10 seats, and 5 sites. The allowance counts model replies, not conversations; at the default 3 replies per conversation, 1,000 replies represent roughly 333 conversations.

For the same 333 resolved conversations, Fin’s published comparison rate of $0.99 per resolution gives an illustrative total of $329.67, before seats. That is not an exact like-for-like quote: Fin counts resolved conversations, while MivoChat counts AI replies, and some technical threads need more than three responses. Starter and Pro are flat prices with no per-resolution fee. Paid checkout remains in development; start with the 14-day trial without a card.

Put chat beside the reference pages

Install the widget on the documentation host, ideally in the shared layout used by reference pages and guides. For Docusaurus, use the supported site-wide integration point so the widget follows documentation navigation without editing every MDX file. For Next.js docs, add the script in the layout with next/script. For a simple static site, add the script to the common page template: <script src="https://mivochat.app/widget.js" data-site="YOUR_SITE_KEY" async></script>.

Start with a narrow source set: authentication, the currently supported API version, common errors, and one troubleshooting guide. Ask the actual questions listed above, including a version-specific request and an undocumented behavior. The answer should match the docs; the undocumented case should hand off. If you cannot identify the relevant public documentation yourself, improve the documentation before inviting customers to rely on that answer.

Do not make the bot an API debugger

MivoChat cannot inspect a customer’s request logs, private repository, API key, or account configuration. It cannot run a curl command, rotate a credential, retry a webhook, or change a subscription. Avoid language implying a generated code response has been tested against a live account.

Use a handoff for reports that require a trace ID, customer-specific usage data, an incident timeline, or interpretation of an unfamiliar payload. The assistant can say it does not know and carry the conversation to your team. Publish a safe process for sharing diagnostics, and tell customers never to paste secrets into public chat.