What a BYOK AI chatbot is, in plain terms
Every AI chatbot has two costs: the platform (the widget, the dashboard, the inbox, the hosting) and the AI itself (the tokens a language model reads and writes for each message). Most website chatbot builders bundle the two. You buy a plan, and the vendor pays a model provider behind the scenes, then meters you with message credits, “AI conversations” or a fee per resolved ticket.
A BYOK chatbot pulls the two apart. You open an account with the model provider you prefer, create an API key, and add it to the chatbot platform. The platform sends each visitor message to the provider with your key, so the provider’s invoice for the AI part comes to you, itemised at the provider’s published rates. The platform’s plan only pays for the platform.
| Bundled AI (typical) | BYOK (SiteTrellis) | |
|---|---|---|
| Who holds the AI account | The chatbot vendor | You |
| How the AI is billed | Inside the plan: credits, messages or resolutions | By your provider, per token, at its rates |
| Choosing a model | The vendor’s menu, often tied to a credit multiplier | Any model your key can call |
| What the plan measures | Often AI usage itself | Platform limits: conversations, sites, agents, seats |
| What you must do | Nothing beyond paying | Create a key, keep credit at the provider, set a spend limit |
The last row is the honest trade-off. BYOK gives you control and visibility, and in exchange you run one more account. If you want to pay a single invoice and never think about tokens, a bundled product may suit you better.
Why website owners choose a BYOK chatbot
- A cost you can read. Your provider’s dashboard shows exactly what each model cost. SiteTrellis also estimates cost per agent and model in its analytics, and the provider’s invoice is always the final word.
- One predictable unit. On SiteTrellis a conversation is one visitor session with at least one visitor message, whichever model answers. There are no per-resolution or per-seat AI fees on top.
- The right model for the job. Use a small, inexpensive model for simple questions and a stronger one for the hard ones, or mix them across fallback steps.
- No lock-in on the model. When a better or cheaper model ships, you change a setting instead of switching products.
None of this makes a BYOK chatbot automatically cheaper: at tiny volumes a bundled free tier might cost nothing. The case is transparency and control, and whether it saves money depends on your volume and model, which is what the cost calculator is for.
How the SiteTrellis BYOK setup works
Create a key at your AI provider
Open the provider’s console, create an API key for SiteTrellis and, if you can, set a monthly spend limit. We have a step-by-step guide for each provider, for example OpenAI and Anthropic.
Add the key to SiteTrellis
Paste it under AI providers in the dashboard. It is encrypted at rest, tested once, and afterwards shown only as a short hint. It is never sent back to your browser.
Build the agent
Write the prompt, paste the facts the agent should answer from (FAQs, policies, product details), and choose a model for each step of the chain. Try it in the playground, then publish.
Add a fallback
Put a second model, ideally from a different provider, behind the first one so a rate limit or outage does not leave visitors with a dead chat.
Paste one script tag
Add the snippet before the closing body tag of your site, or follow the guide for WordPress, Shopify, Wix or Squarespace.
What you pay: the platform plus your provider
There are two lines on your bill, and only one of them is ours. The plans below cover the platform. The AI is billed by your provider.
| Plan | Price per month | Conversations per month | Websites | Agents | Fallback steps per agent |
|---|---|---|---|---|---|
| Free | $0 | 100 | 1 | 1 | 2 |
| Starter | $19 | 1,000 | 1 | 2 | 3 |
| Pro | $59 | 5,000 | 5 | 10 | 5 |
| Agency | $149 | 20,000 | 25 | 50 | 5 |
Plans cover the platform only. Your AI provider bills the tokens. Full plan comparison.
For the AI part, the cost depends on the model, the length of your prompt and knowledge, and how long conversations run. As a worked example, with 6 visitor messages per conversation, 1,500 tokens of prompt and knowledge sent every turn, and 180-token replies, our model catalog gives an estimate of roughly $1.79 per 1,000 conversations on Claude Haiku 5.5 and roughly $35.88 per 1,000 conversations on Claude Sonnet 5.5. Those are list-price estimates from our catalog, not quotes: your provider’s bill is authoritative, and real conversations vary a lot.
Set a monthly spend limit at your provider before you publish. It is the hard safety net behind SiteTrellis’s own rate limits and message caps.
To try your own numbers, use the AI chatbot cost calculator or read how much an AI chatbot really costs.
Which provider’s key should you bring?
SiteTrellis supports five providers, and you can mix them in one agent. Each page below covers what to pick, how to set up the key, and what it tends to cost.
Claude (Anthropic)
Strong instruction following and careful answers. Prepaid credit at the Anthropic Console.
OpenAI
The most familiar API. Note that API billing is separate from a ChatGPT subscription.
Google Gemini
Fast, low-cost models and a free tier with low limits.
DeepSeek
Low list prices; read the data-location terms before you use it.
OpenRouter
One key that reaches models from many labs, handy for fallbacks.
Not sure where to start? How to choose an LLM for website chat covers cost, speed and quality.
Why a BYOK chatbot should have more than one key
When you bring your own key, the provider’s limits become your limits: a rate limit, an empty balance or an outage at that one provider stops your chat. That is why every SiteTrellis agent has an ordered fallback chain. If a step fails with a network error, a timeout, a 429, a 5xx response or an invalid key, the next step answers, and the conversation history is carried over so tool calls survive the switch.
The number of steps depends on your plan: Free 2, Starter 3, Pro 5. A good pattern is a fast, low-cost model first and a different provider second. See why one provider is a single point of failure.
Keys, data and safety
- Keys are encrypted at rest with libsodium, shown masked after saving, and never returned to the browser or written to logs.
- Per-IP, per-session and per-site rate limits and message-length caps sit in front of your key, so a bot filling your chat cannot drain your balance.
- Your provider receives the text needed to write each reply, under your agreement with that provider. Mention it in your privacy policy.
More in Security and privacy. Our terms, privacy policy and data processing agreement are on the Legal page.
What this is not (yet)
Honest limits: knowledge is text you paste. There is no website or PDF crawler yet, replies arrive whole instead of streaming word by word, and there is no voice, WhatsApp or other messaging channel. Those are common in bundled chatbot builders, so check you do not need them.
You also need a funded provider account. If you would rather not manage one, a bundled chatbot is a fair choice. If you do want one, BYOKChat vs SiteTrellis compares us with the closest BYOK tool, and our alternatives pages cover the managed products.