Free AI for testing

You do not need a paid plan to try AI insights. Pick one of the routes below, wire the model into your server assembly, then enable a profile for a test project.

Use test data only

Free tiers can log prompts or use them to improve models, and free models can change or disappear. Evidence includes story names, logs and (with vision) screenshots. Do not point a free-tier profile at a private project. Check each provider’s current terms before use.

Local model (no data leaves your machine)

Run a local server such as Ollama, which exposes an OpenAI-compatible API, and connect through the OpenAI-compatible provider:

import { createOpenAICompatible } from "@ai-sdk/openai-compatible";
import { createAi } from "@storyshelf/ai";

const local = createOpenAICompatible({ name: "ollama", baseURL: "http://localhost:11434/v1" });
const ai = createAi({
  profiles: { default: { defaultModel: local("<model-id you pulled>") } },
  budget: { perProjectCallsPerHour: 20 },
});

Small local models are slow and may need a longer timeoutMs ({ model, timeoutMs: 180_000 }). Skip screenshots by leaving the model text-only (the default).

Google AI Studio

Create an API key in AI Studio, then use the Google provider with a model that is available on your key’s free tier:

import { createGoogleGenerativeAI } from "@ai-sdk/google";

const google = createGoogleGenerativeAI({ apiKey: process.env.GOOGLE_API_KEY });
const ai = createAi({
  profiles: { default: { defaultModel: google("<free-tier model id>") } },
  budget: { dailyTokens: 200_000, perProjectCallsPerHour: 10 },
});

OpenRouter

OpenRouter offers some models at no cost and many at low cost behind one API key:

import { createOpenRouter } from "@openrouter/ai-sdk-provider";

const openrouter = createOpenRouter({ apiKey: process.env.OPENROUTER_API_KEY });
const ai = createAi({
  profiles: { default: { defaultModel: openrouter("<model id>") } },
  budget: { dailyTokens: 200_000, perProjectCallsPerHour: 10 },
});

Cap reasoning on small models

Free-tier and small reasoning models can spend the whole output cap thinking. Lower the effort for the slot with providerOptions (see Setup), for example { model, providerOptions: { google: { thinkingConfig: { thinkingBudget: 0 } } } }; check your provider’s current option names.

Tips

• Keep the budget small so a test run cannot burn through a free quota.
• Start text-only; add a vision model only when you want screenshots analyzed.
• Quality varies a lot between models. If answers are rejected as invalid, try a model with reliable JSON output.
• Remove the profile (Project settings → AI → Off) when you finish testing.