Never sleeps.
Never rented.

Yes. You pay once and run the software on your own infrastructure for as long as you like. There is no per-minute charge, no per-seat charge, and no renewal required to keep it running. You do pay Google for Gemini and Twilio for telephony, but you pay them directly at their rates, with nothing added by us.

Nothing stops. The version you own keeps running exactly as it did, forever, offline. What ends is your access to new releases and support. You can renew for updates if you want them, or stay where you are indefinitely. A licence carries the highest major version it activates, so upgrading to a future major version is a decision you make when it suits you.

No, and we will not pretend otherwise. Your licence verifies offline using a signature your deployment checks locally, so it never contacts us. That means we genuinely cannot reach into a running deployment. Revoking a licence stops future issuance, image pulls, and support; it cannot stop software that is already running. That is the honest cost of never phoning home, and we think it is worth it.

Onto your own infrastructure and nowhere else. Each deployment is single tenant: your own database, your own credentials, your own phone numbers. Conversations, transcripts, and knowledge base documents never touch shared infrastructure, and never reach us. SQLite is the default and needs nothing extra, or you can point it at your own Postgres.

For the live voice and video conversation, yes, and that is a technical boundary rather than a preference: no other provider currently offers an equivalent realtime bidirectional audio API. Everything else is open. An Orchestrator, the layer that carries out multi-step delegated tasks, runs on Anthropic, OpenAI, a locally hosted model, or any of a hundred others. An agent on a live phone call can hand a task to a Claude-backed Orchestrator mid-conversation and speak the answer back in its own voice.

Only for phone calls and WhatsApp. The embeddable website widget, in both text and voice, needs nothing beyond a Gemini key. Plenty of deployments start as a widget on a marketing site and add a phone number later.

By capacity rather than throughput: how many conversations run at once, how many agents you configure, how many people administer it. Those are the things a self-hosted deployment can measure honestly at a single instant. Metering minutes would require your deployment to report usage back to us, which would break the one guarantee that makes the product what it is.

Your plan states a number: one on Basic, two on Pro, unmetered on Enterprise. Where you scale out, the software enforces it. Instances of one deployment share a database, so they count each other there, and an instance beyond your number keeps its dashboard and finishes the calls it already has but starts no new ones. That needs no phone home and no hardware lock: the count never leaves your own database, and an air-gapped deployment behaves exactly like a connected one. Where we cannot enforce it is across wholly separate deployments with separate databases, and we will not add a phone-home or a machine lock to change that, so there the number stays a term of the licence. Worth knowing: capacity limits like concurrent calls are counted against the database, so instances sharing one share that ceiling, while a separate deployment gets its own.

Two. A demo mode that runs forever with one agent and a website widget, which is enough to put a working assistant on a site and see how it behaves. And a 14 day trial with every feature unlocked, self-serve, no card, for when you want to prove it on real work.

At low volume. If you are running a few hundred conversation minutes a month and expect to stay there, a metered platform will cost you less than a licence for a long time, possibly forever. The cost calculator on the pricing page will tell you so plainly rather than talking you out of it. The economics turn in our favour with volume and with time, not immediately.

It keeps working. A verification failure, whether from a clock drift, a mistyped key, or an expired trial, opens a 14 day grace period during which the deployment runs exactly as before, logs loudly, and shows an escalating banner in the dashboard. Nothing about a licensing problem should be able to take a phone line down overnight.

Write a skill. A knowledge base document is searched for the passage that answers a question; a skill is a short procedure you write, with a description of when it applies, and the whole thing loads at once when it matches. Half a refund policy is worse than none, which is why skills are not chunked. Orchestrators keep their own library separately. Demo mode allows five, every paid plan is unmetered.

Yes, on Enterprise. Federation registers other deployments as parent, child, or peer, and lets them delegate work to each other. The part worth knowing is what stops it going wrong: every delegation carries a hop count and a trace across the instance boundary, so a loop between two deployments is refused rather than merely logged, and that guard is on for everyone regardless of plan. Inbound work can be restricted to keys you have bound to a link, rate limited per link, and a link that keeps failing disables itself and says so.

Exactly what you scoped its key for. AiFlow runs as an MCP server, and access is granted one scope per capability: placing calls, searching the knowledge base, reading transcripts, listing agents, recording events, looking up a contact, calling a custom tool, delegating to an Orchestrator. There is deliberately no wildcard scope, so a key issued for reading transcripts cannot place a call. Every call an outside client makes is written to the audit log.

Salesforce and HubSpot have native connectors: OAuth, tokens encrypted at rest, and tools for upserting a contact, creating a ticket, and logging call activity. Beyond those, AiFlow speaks MCP in both directions. A catalog of vendor-hosted servers, Notion and Linear and Stripe and Sentry among them, connects in one click, and anything not in it connects by pasting a URL, because AiFlow implements the protocol rather than one connector per service. It also runs as an MCP server itself, so a client such as Zendesk or Claude Desktop can call into it. There is no native Zendesk connector, and that is the honest shape of that integration: it runs the other way round.

Two answers depending on what it is. If it is an integration, custom tools and MCP connections let you extend an agent without us writing anything: describe the action, point it at an endpoint you control. If it is deeper than that, MeridFlow builds it. Custom tools, a bespoke MCP server over your internal systems, integration into an architecture you already have, or a managed and hardened deployment.

That depends on your hardware and your Gemini and Twilio quotas rather than on the software, and each plan sets a licensed ceiling. Running more than one instance behind a load balancer is safe and is what the instance number on your plan is for: they share a database and count each other against it. What is not built is coordinated throughput scaling, so a second instance answers more sessions but does not make the outbound call queue itself go faster, and rate limits are per instance rather than pooled. If you need that shape of scale it is an engagement with us rather than a checkbox.

On Enterprise, yes. One setting changes the name, logo, and colours across the dashboard, the login page, and every embedded widget, with no fork and no rebuild. Enterprise also carries resale rights, which is what agencies running it for their own clients need.