Guide · Beta
Bring Your Own Key
Run Sonny's AI on your own OpenAI or Anthropic account — or a self-hosted, OpenAI-compatible server — with unlimited AI conversations. A one-time $249 unlock, no extra subscription. Here's what you get and how to set it up.
Want the full pitch and cost comparison? See the Bring Your Own Key overview.
What the unlock gets you
Unlimited AI conversations
Monthly conversation allowances disappear. Your provider bills you per token, and a typical support conversation costs around a cent on an efficient model like Claude Haiku.
Your provider, your model
Connect OpenAI, Anthropic (Claude), or any OpenAI-compatible endpoint — including a self-hosted Ollama or vLLM server. Choose from curated models or enter any model ID.
Control where AI processing happens
Every customer-facing response is generated on your account, under your provider agreement. Ideal for custom server environments and data-residency requirements.
Keys stay secret
Your API key is encrypted at rest and never displayed again — only a hint like "sk-a…4f2a" so you know which key is connected.
All AI features included
The auto-responder, AI reply drafts, and polish all run through your key. Behavior, confidence thresholds, and office-hours schedules work exactly as before.
Knowledge base stays managed
Sonny keeps crawling your website, processing documents, and running embeddings and retrieval at no extra cost. You only bring the part that writes the answers.
Setting it up
Unlock BYOK ($249, one-time)
On the AI Copilot page, find "Bring your own key" and click Unlock. It's a single purchase per workspace — not a subscription — and only the workspace owner can buy it.
Connect a provider
Click "Set up your key", pick OpenAI, Anthropic, or OpenAI-compatible, and paste your API key. Self-hosted endpoints also need their URL (it must be reachable from the internet — include the /v1 path if your server uses one).
Test the connection
Sonny runs two live checks: can it reach your provider and authenticate, and can the model make tool calls (the auto-responder needs those to search your knowledge base and hand off to humans). You can't enable a connection that hasn't passed.
Save & enable
From that moment, every AI response runs on your account. The AI Copilot page shows "Running on your key" and your usage stops counting against any plan allowance.
A note on cost
A typical AI support conversation uses about a cent of API tokens on Claude Haiku. At 5,000 conversations a month that's roughly $50 in token costs on your own key, compared with the $199/mo AI Scale plan — so heavy users usually recoup the $249 unlock within the first month or two. Your exact costs depend on the model you pick and how long your conversations run.
Good to know
What happens if my key stops working?
Customers are never left hanging. If your provider rejects a request — an expired key, a rate limit, a server outage — the auto-responder immediately routes those conversations to your team, and workspace owners get a notification with the error. Fix the key (or toggle back to Sonny AI) and everything resumes.
Can I switch back to Sonny AI?
Anytime. The card has an on/off toggle — turn it off and Sonny AI takes over instantly (with an active AI plan). Your key and settings are kept so you can switch back with one click. You can even keep an AI plan and BYOK active together; whichever is enabled handles responses.
Which models should I use?
Sonny is tuned for Claude Haiku-class models and above — that's what Sonny AI itself runs on. On OpenAI, a mini-class model is a great fit. Smaller self-hosted models can work, but expect lower answer quality and less reliable escalation; the built-in tool-calling test will tell you if a model isn't up to the job.
Do knowledge base pages still have a cap?
Yes — BYOK includes a 5,000-page knowledge cap, because crawling and embedding stay on Sonny's infrastructure at our cost. If you also have an AI plan with a higher cap (like Scale's 20,000), you keep the higher one.
Is my API key safe?
Keys are encrypted at rest with AES-256-GCM, never returned by our API after saving, and only decrypted server-side at the moment a request is made to your provider. Self-hosted endpoint URLs are validated so requests can only go to your server.