Function calling over the real API
Every action the assistant can take is a real endpoint on the engine, invoked with your identity. No shadow implementation, no scripted shortcuts, no capability that exists only for the chatbot.
AI assistant
The Zinn Digital® assistant has function-calling access to the same action catalogue that powers the dashboard, so it can reset SSL, add disk, purge cache or open a ticket on your behalf — always within your own permissions. It is grounded in your knowledge base and your resolved tickets, and it speaks every one of our 58 languages.
Most hosting "AI" is a search box with a friendlier tone. Ours is wired into the platform API through function calling, so a request becomes an executed action with a confirmed result.
The assistant sits on top of the same action catalogue the dashboard, the public API, the CLI and our MCP server all use. There is no separate AI code path to drift out of sync: when we ship a new capability to the platform, the assistant can use it, gated by exactly the same permission checks as a human clicking the button.
Ask it to reset a certificate, add disk to a site, purge the CDN, open a support ticket or walk you through a screen with a guided tour, and it carries the request out and reports back — rather than telling you where the setting lives.
Every action the assistant can take is a real endpoint on the engine, invoked with your identity. No shadow implementation, no scripted shortcuts, no capability that exists only for the chatbot.
Access resolves through the same role and permission catalogue as the rest of the platform — owner, billing, dev and read-only customer roles, mapped to granular module.action keys. The assistant can never exceed what your membership allows.
Every read and write is scoped to your organisation and enforced by Postgres row-level security, the same isolation the API itself runs under. There is no cross-tenant query path for the assistant to take.
Actions taken through the assistant are audit-logged with actor, action and target, so there is a record of what was done, by whom, and on whose behalf.
An assistant that hallucinates hosting advice is worse than no assistant. Ours answers from a retrieval layer built on real, tenant-scoped content.
We run a core RAG layer on Postgres with pgvector, holding embedded knowledge-base articles and resolved support tickets under row-level security. It uses multilingual embeddings, so a question asked in Arabic retrieves an answer written in English without an intermediate translation step degrading the match.
Every ticket that reaches resolution is embedded back into that index. The assistant improves through retrieval rather than fine-tuning — the knowledge grows continuously, stays scoped to the right tenant, and remains transactionally consistent with the knowledge base it came from.
When you start to get help, the assistant does not open a blank form. It runs a live diagnostic on the affected site first, in your language.
It pulls from our observability signals and runs live checks in the same pass: uptime, HTTP and PHP errors, SSL and DNS state, resource limits and LVE faults, malware flags from the scanner, plugin and theme conflicts, database health and cache state. Within a few seconds it has a picture of the site that a human agent would otherwise spend the first reply gathering.
From there it takes one of three routes — fix it, teach you to fix it, or hand it to a human who already knows everything it found.
A structured check across uptime, errors, SSL/DNS, resource limits, malware flags, plugin and theme conflicts, database health and cache — combining stored observability data with live probes against the site.
Where the fix is safe, reversible and inside your entitlement, it applies it through the action catalogue — then re-checks the site and confirms whether the symptom has actually gone.
For anything outside its guardrails it gives step-by-step instructions grounded in the knowledge base, written for your situation rather than a generic article.
If a human is still needed, the ticket opens pre-filled with the diagnostics, what was attempted and the outcome — so the agent starts where the assistant stopped, not from scratch.
The auto-fix set is deliberately conservative: reversible operations, inside your existing entitlement, always followed by a re-check.
Anything beyond that set — destructive operations, spend, or changes with blast radius — needs your explicit confirmation or falls to a human. That boundary is the reason the assistant is safe to give real access to.
The assistant, the diagnostics it reports and the ticket it escalates all run in the customer's language. This is not an English product with a translation bolted on the front.
Support runs a translate-both-ways flow: you write in any of our 58 languages, we store your original and machine-translate it to English for the agent, then translate the reply back into your language on send. The assistant works the same way, and because the retrieval index uses multilingual embeddings, the knowledge behind it is available in every language too.
Translation sits behind a swappable provider adapter, so no single vendor owns a part of the platform your support experience depends on.
One action catalogue, several front doors. What the assistant can do in the dashboard, your own tools can do too.
Every model call — assistant, triage, translation, automations — routes through our AI gateway, which applies prompt and result caching by default and gives us provider fallbacks and rate limiting. Model identifiers are configuration data in a registry, never hardcoded, so we move to a better model without a platform release.
Present throughout the customer dashboard alongside notices and guided tours, so help arrives on the screen where the problem is.
Our MCP server exposes the same catalogue to any MCP-capable agent — Claude Code, Cursor, ChatGPT, Claude Desktop — with OAuth 2.1, org and RBAC-scoped, per-tool revocable tokens.
Recipes that act directly on your sites — auto-blog, image generation, alt text, SEO rewrites, translation — run as durable scheduled workflows using your own provider keys, with per-customer spend caps, optional human approval before publish and action logging.
Support agents run their own AI agent over a broader catalogue, bounded by their staff RBAC and audit-logged, with sensitive actions requiring confirmation. Your escalated ticket lands with context on both sides.
Its auto-fix set is restricted to safe, reversible operations inside your existing entitlement — cache clears, PHP worker restarts, SSL repair, memory within plan limits, table repair, disabling a fatally-erroring plugin — and every fix is followed by a re-check. Anything destructive, anything that costs money and anything outside those guardrails requires your explicit confirmation or goes to a human. That narrows the blast radius substantially, but no automated remediation is risk-free, which is why every action is audit-logged and daily backups sit underneath it.
Every query the assistant makes runs under the same Postgres row-level security as the rest of the platform, scoped to your organisation, and the retrieval index it reads from is tenant-scoped in the same way. There is no cross-tenant query path in the platform for it to use, and the public marketing and help-centre search runs on a separate, non-tenant index.
Exactly what your role permits, and nothing more. Permissions resolve through the platform's granular module.action catalogue against your membership — so a read-only or billing role's assistant can read and advise where those roles can, but cannot invoke actions those roles are not granted.
Two ways. It reads live platform state and observability signals for the site in question — uptime, errors, SSL and DNS, resource limits, malware flags, database and cache health. And it retrieves from an index of knowledge-base articles plus your organisation's own resolved tickets, embedded multilingually and scoped by row-level security.
It gives you step-by-step guidance grounded in the knowledge base, and if a human is still needed it opens a ticket pre-filled with the diagnostics it ran, the fixes it attempted and the results. Support then works in your language: your message is stored in the original and translated to English for the agent, and the reply is translated back on send.
Yes — all 58 languages we support, throughout. The diagnostic run, the explanations, the guided steps and the ticket thread all operate in the customer's language, and the retrieval layer uses multilingual embeddings so the underlying knowledge is reachable regardless of the language the question is asked in.
Yes. The assistant is one interface onto the action catalogue; our MCP server is another. Connect Claude Code, Cursor, ChatGPT, Claude Desktop or any MCP-capable client with OAuth 2.1 and org-scoped, per-tool revocable tokens, and drive the same capabilities from wherever you already work.
AI Pack automations use your own provider keys, held in our credential store, so you pay your model provider directly for what your recipes consume. Per-customer spend caps, rate limits, action logging and optional human approval before publish bound that usage. Plan inclusions are listed on the pricing page.
Start a card-free 7-day trial, break something on purpose, and watch the assistant find it, fix what is safe and explain the rest. 30-day money-back guarantee, free migrations, no lock-in.
Start free