Frequently Asked Questions
Answers to common questions about Arcanflows
What Arcanflows does, what it runs on, what it costs, and where the limits are. If something here is out of date, tell us — every answer is meant to describe the product as it ships today.
Getting started
What the platform is and how quickly you can have something working.
What is Arcanflows, in one paragraph?
A platform for building AI systems that talk to people and do work. You create agents that answer on chat, email or a real phone number; workflows that run on a schedule, a webhook or an agent's decision; forms and data tables to hold the information; apps to put a branded interface on top; and Studio to generate video. It is one product rather than several tools bolted together, which is why context survives when a conversation moves from an agent to a workflow to a human.
Do I need to write code?
No. Agents, workflows, forms, data tables and apps are all built through visual editors, and the Builder will design a whole working system from a plain-language description. Code is available where you want it — JavaScript transforms inside workflow nodes, Python and JavaScript script tools, and a full REST API — but nothing requires it.
How long until I have a working agent?
Minutes for a first agent: pick a model, write a system prompt, optionally upload documents for it to answer from, and publish. A production deployment with workflows, integrations and a human handoff path is realistically a few weeks of work — most of which is deciding what should happen, not building it.
Can I try it without paying?
Yes. The Free plan costs nothing, needs no credit card, and includes the agent builder, workflows and forms with monthly quotas. Paid plans are Pro, Enterprise and Enterprise Premium — the pricing section on the home page always shows current prices and limits, because it reads them from the billing system rather than from hand-written copy.
AI models and providers
Which models you can use, who supplies them, and who pays for the tokens.
Which AI providers can I connect?
OpenAI, Anthropic, Google, xAI, Mistral, Cohere, DeepSeek, Groq, Alibaba and Together are supported as commercial providers, plus any OpenAI-compatible endpoint. You manage them under Commercial Models in the console, where you can also sync a provider's catalogue so new model releases appear without waiting for a platform update.
Can one account use several providers at once?
Yes, and most do. Different agents can run on different models — a cheap fast model for triage, a stronger one for analysis, a local model for anything sensitive — and a single agent can name fallback models to try if its primary is unavailable. Nothing forces you to standardise on one vendor.
How does the platform choose which model handles a request?
You choose, per agent. Where several backends can serve the same model you also pick a routing strategy — priority (with automatic fallback to the next backend), weighted, round robin, least loaded, or random — under Advanced Settings in the agent editor. It is deliberate configuration rather than a hidden heuristic, so you can always explain why a given request went where it did.
Do I use my own API keys, or yours?
Either. Bring your own provider keys and you pay that provider directly at their rates, with the keys held in an encrypted credential vault. Or use the platform's included model access and consume the token quota on your plan. Teams typically start on the included quota and move to their own keys as volume grows.
Which media models are available beyond text?
Image generation and editing, video generation, text-to-speech, speech-to-text, and music and sound effects — from providers including OpenAI, xAI, Google, Runway, Replicate, ElevenLabs and Alibaba, depending on what you connect. These are usable as agent tools, as workflow steps, and as nodes in Studio.
Your own models and infrastructure
Running inference on hardware you control, and why you might.
Can I run models on my own servers?
Yes — this is a first-class part of the platform, not an afterthought. You register inference backends and the platform routes to them exactly as it does to a commercial provider. Supported backend types are Ollama, vLLM, Text Generation Inference, MLX (Apple Silicon), SGLang, llama.cpp, and any custom OpenAI-compatible endpoint, plus separate speech-to-text and text-to-speech backends.
Do I have to set that up by hand?
No. The deployment wizard walks through it: choose the backend type, choose a model from the popular list or name your own, set parameters such as GPU layers and context length, choose where it runs — local Docker, a remote server over SSH, or Kubernetes — and review before it deploys. Afterwards you get health monitoring, GPU and performance metrics, and the ability to pull new Ollama models straight from the console.
Why would I run local models instead of using a commercial API?
Three usual reasons: data that must not leave your infrastructure, cost at high volume, and independence from a vendor's availability and pricing. The practical answer is usually a mix — sensitive workloads on local models, everything else on commercial APIs — which the platform supports directly because both are just backends.
Does the rest of the platform work with local infrastructure too?
Yes, including voice. Studio's voiceover and transcription nodes fall back automatically to your local voice backends when no commercial provider is configured, and phone agents can run speech-to-text and text-to-speech on your own servers. Local runs still count against your plan quota, because the platform is doing the orchestration either way.
Can I host the whole platform myself?
Yes. Arcanflows runs as Docker containers and is designed for self-hosted deployment — single server, multiple servers, or Kubernetes. You keep your data, your models and your infrastructure.
Building things
The Builder, workflows, data and apps.
What does the Builder actually do?
You describe what you need in a sentence or two. It designs the whole system — data tables and their relations, forms, agents, workflow steps and a dashboard — and shows you every piece for review before anything is created. Everything it produces is ordinary platform objects you can edit afterwards, so it is a starting point rather than a black box.
What can a workflow do?
Over twenty node types: call agents and tools, branch on conditions, loop over collections, transform data with JavaScript, call HTTP APIs, read and write data tables, send email, wait, and run sub-workflows. Triggers include schedules with real timezone handling, webhooks, form submissions, chat, and other workflows. Retries and error handling are built in.
Where does my data live inside the platform?
In data tables — a spreadsheet-like store with typed columns, views, filters, imports, and relations between tables. Forms write into them, workflows read and update them, agents can query them, and apps display them. There is also a document store for the files agents answer from.
Can I give my team a real interface, not just a chat box?
Yes — apps are branded internal interfaces built from your tables, forms and agents, published to the people you choose. Themes and brand kits mean apps, forms and embedded chat carry your colours and logo rather than ours.
Conversations with your customers
Chat, embedding, phone, and what happens when the AI should stop.
Where can an agent talk to people?
In the console, on your own site through an embeddable widget added with one script tag, over email, and on a real phone number. It is the same agent with the same knowledge in each channel — you are not rebuilding it per surface.
Can an agent answer the phone?
Yes. Agents can be given an extension and receive real calls over SIP, with speech-to-text and text-to-speech on either commercial or your own local voice backends. Every call is recorded in the Calls page afterwards with its transcript, and voice minutes are metered against your plan.
What happens when the AI should not be answering?
The conversation moves to a person without restarting. A customer can ask for a human with a phrase you choose — which escalates before the model even replies — or the agent can hand over on low confidence or a rule you set. Your operator arrives with the full history, the AI stays quiet while a human owns the conversation, and the operator can transfer to a colleague with a note or resolve and let the AI resume.
Can several agents work together?
Yes. Collaborations run agents in sequential, router, parallel or debate patterns — one agent routes to specialists, several answer in parallel and a final one synthesises, or two argue a question through. Agents can also call each other as tools.
Studio — generating video
The node canvas for making films with AI, and what it can realistically produce.
What is Studio?
A node canvas for creating video. Each node produces something — a prompt, an image, a clip, a voiceover, music — and connecting nodes carries one output into the next. The usual chain is prompt to image, image to video, then several clips plus a voiceover assembled into a finished film with cuts or crossfades.
Video models only generate a few seconds. How do you get a longer film?
By chaining. A video node can continue from the connected clip's last frame, so each new clip picks up exactly where the previous one ended, and the assemble node joins them into one film. That is how you get past a single generation's length limit — the constraint becomes your patience and your quota, not the model.
Can I put a consistent character on screen?
Yes. The avatar node takes either a preset character from a curated gallery or your own uploaded photo or written description. With an Alibaba (DashScope) video backend the character image is used as an identity reference on every clip, so the character does not drift between scenes and clips render in parallel; otherwise the platform chains clips through Veo, repeating the character description each time.
Can characters speak, with the mouth actually matching?
With a Google Veo backend configured, presenter avatars speak with natively synced mouth movements — dialogue you write in quotes is spoken by the character. Without one it falls back automatically to motion plus a voiceover track, so presenters still work everywhere; only the lip sync differs.
Do my agents have any role in Studio?
They can direct. An agent node runs one of your agents as a creative collaborator — an art director that turns one line into a full visual brief, a scriptwriter, a shot-list writer — and its answer flows straight into the image, video and voiceover nodes. Several agents with different roles can contribute to a single film.
Automation and integrations
Connecting the platform to everything else you run.
What can Arcanflows connect to?
Any REST API, plus one-click connectors for common services such as GitHub, HubSpot, Jira, Notion and Slack. Authentication covers OAuth 2.0, API keys, bearer tokens, basic auth and custom schemes, with credentials held in an encrypted vault. Webhooks work in both directions, and agents can query databases through tools.
Can things run on a schedule without anyone present?
Yes. Scheduled tasks run agents and workflows on a recurring schedule with correct timezone handling, and deliver the result where you want it — email, a data table, a chat channel, or a webhook. A daily report that assembles itself and lands in an inbox is a few minutes of setup.
Is there an API and a command-line tool?
Both. The REST API covers the platform, and agents can be called from your own applications with a scoped API key. The ARCANFLOWS CLI brings agents to your terminal — chat, run workflows, query data tables, analyse repositories — on macOS, Linux and Windows.
API documentation · Download the CLI
Can I extend the platform with my own code?
Yes. Script tools run Python or JavaScript in a sandbox, workflow transform nodes run JavaScript inline, and custom REST tools wrap any API you already have. Anything you build this way is available to agents exactly like a built-in tool.
Security, governance and compliance
What protects the platform, and what it does about prompt injection.
How is access controlled?
Five built-in roles — super admin, tenant admin, agent developer, agent user and viewer — whose individual permissions each tenant can customise. Multi-factor authentication is available, tenants are isolated from one another, and activity is recorded in an audit log.
How are credentials and keys stored?
In an encrypted credential vault, using AES-256-GCM. Secrets are referenced by workflows and tools rather than pasted into them, and are never shown back to the model.
What stops someone talking an agent into misbehaving?
A platform-wide guardrail layer screens both what goes into the model and what comes out, on streaming and non-streaming paths alike. Content the agent retrieves — tool results, scraped pages, documents — is screened separately and neutralised rather than blocked, because that is where real attacks begin: text an attacker planted somewhere the agent will read it. Rules update without a deploy, and blocked or neutralised events are logged and can raise alerts.
Do you train models on my data?
No. Your content is not used to train our models. When you route a request to a third-party provider, that provider's terms govern what they do with it — which is one of the reasons the platform makes local models straightforward.
Is the platform compliant with HIPAA, GDPR or similar?
Compliance is a property of a deployment, not of software, so the honest answer is that the platform provides the controls — role-based access, MFA, audit trails, encrypted secrets, tenant isolation, retention settings, self-hosting and local models — and the compliance of any given installation depends on how and where you run it and on the agreements you hold with your providers. Talk to us about your specific requirement rather than relying on a badge.
Plans, quotas and billing
What you are charged for and how usage is measured.
What are the plans?
Free, Pro, Enterprise and Enterprise Premium. Current prices, agent and user limits and monthly quotas are shown in the pricing section on the home page, which reads live from the billing system — so it is accurate even when this page is not.
What exactly is metered?
Eight buckets: model tokens, images, video, audio generation, text-to-speech, speech-to-text, voice minutes and storage. Every one is visible in your dashboard as used-against-limit, so nothing is a surprise at the end of the month, and usage from workflows, scheduled tasks and phone calls is counted the same as usage from a chat window.
What happens when I hit a limit?
The affected capability stops rather than silently billing you more, and the dashboard shows which bucket ran out. Upgrading or connecting your own provider keys both restore service — with your own keys, model usage is billed by that provider instead of consuming platform quota.
Still have a question?
Ask us directly — a person answers. If you already have an account, the Platform Assistant inside the console can answer most product questions instantly, because it reads the same documentation this page is written from.
Email [email protected], or see the full FAQ at arcanflows.com/faq.