Persona Lab: one chat, five LLM providers, no usage cap.

What it is

Persona Lab is AI focus-group software for marketers, product managers, and content strategists. Instead of recruiting humans and paying for a traditional focus group, you spin up a roomful of AI personas with distinct backgrounds, voices, and biases, then drop a piece of copy or a product idea into the middle and watch them argue about it. It is not a single-persona chat. It is a multi-persona group conversation where the personas talk to each other, raise objections, and surface what an audience would actually notice.

The product is in production at personalabai.com. Three tiers: free, $9 a month, $49 a month. Stripe handles billing. Google handles sign-in. The user's AI API keys are encrypted per-user with AES-256 and the user pays whichever LLM provider they choose directly, at the provider's published rate. No markup, no per-message fees.

This is a 2 Acre Studios product, not a client engagement. It is the thing we point to when prospects ask whether we can actually build an AI SaaS end-to-end, rather than a slide deck claiming we could.

The engineering problem worth talking about

Routing one user-facing chat across five different model providers is the part nobody warns you about. The marketing surface says "supports OpenAI, Anthropic, Groq, Mistral, Ollama" and ships in a paragraph. The reality lives in four or five thousand lines of glue.

Every provider has a different opinion about what a "message" is, what a "tool call" returns, what counts as an error, and what "streaming" means at the byte level.

The five places they disagree

• Message format. OpenAI wants role/content. Anthropic wants the system prompt as a top-level field, not inside messages. Groq mostly follows OpenAI but quietly drops certain parameter names. Mistral has its own variant. Ollama follows whatever the underlying model was trained on, which means it changes per model.
• Streaming. Server-sent events, but with different chunk schemas. OpenAI emits delta diffs. Anthropic emits typed events (message_start, content_block_delta, etc.). Some providers send keepalives that look like content if you are not careful.
• Errors. Rate-limit signals are 429 on most, but the shape of the retry payload differs. Some include retry-after; some include a JSON body with a wait field; some return 200 with an error inside the streaming body.
• Context limits. Token accounting is per-provider and per-model. A 32k window on one provider is not the same 32k on another after you account for system prompt overhead, tool definitions, and the provider's internal scaffolding.
• Cost shape. Per-token in/out with different cache discount rules. Some providers charge for cached prefix, some do not. Some charge for tool-call tokens, some absorb them.

How we solved it

A normalized internal message format, one adapter per provider, and a router that picks the adapter at runtime based on the user's stored key. Each adapter is responsible for translating into and out of the internal format and emitting a uniform stream of events to the front end. The front end has no idea which provider it is talking to. The model picker is a drop-down; the routing is invisible.

The win of this approach is not "we support every model." Every other product claims that. The win is that the user does not pay us a usage markup. They pay us a flat $9 a month for the persona framework, the group-chat machinery, the templates, the encryption layer, and the dashboard. The model calls go straight from their browser session to their own provider account.

The pricing call

The competitive comparison is not subtle. OpinioAI charges $99 a month with credit-based caps. Wynter starts at $798 a month. Persona Lab charges $9 a month flat with no message cap, because the architecture lets the user bring their own keys. Bring Your Own Key is not a marketing tagline here; it is the load-bearing pricing decision.

This means a Pro user paying $9 a month gets unlimited AI conversations. The only cost beyond that is what they pay their AI provider directly. The user is in control of which model, which provider, and what they spend on inference. We are in control of the persona framework, the group-chat machinery, the templates, and the data plane.

Most SaaS pricing models force the architecture. We let the architecture pick the pricing model. It rules out a lot of growth-hacker tricks. It also rules out the failure mode where the user opens the app, gets a usage warning, and bounces.

What this means for clients

When 2 Acre Studios ships AI for a client, this is the kind of product we are shipping. Not a wrapper around one LLM. Not a chat box bolted onto someone else's API. The thing we are pointing to is in production and operating. The same orchestration patterns, the same encryption pattern for stored credentials, the same provider-agnostic adapter layer, the same flat-pricing logic.

If you are thinking about building an AI feature or an AI product and the conversation keeps stalling on "is this even possible at our budget," Persona Lab is the answer. We have done it. You can use it. You can also pay us to do the same thing for your business.