Allow user to select provider

CLAUDE.md

@@ -6,9 +6,9 @@ Frontend on :3500, backend on :3501, Mastra Studio on :4111, Convex dashboard on
 
 ## Setup
 
-1. Copy `.env.example` to `.env` and fill in your keys:
+1. Local mode creates `.env` automatically and collects TinyFish + LLM provider credentials in the setup UI. Production still uses env keys:
    - `TINYFISH_API_KEY` — from https://agent.tinyfish.ai/api-keys?utm_source=github&utm_medium=organic&utm_campaign=bigset-developer-2026q2
-   - `OPENROUTER_API_KEY` — from https://openrouter.ai/settings/keys
+   - `OPENROUTER_API_KEY` — production default LLM provider key
    - `NEXT_PUBLIC_CLERK_PUBLISHABLE_KEY` — from Clerk API Keys
    - `CLERK_SECRET_KEY` — from Clerk API Keys
    - `CLERK_JWT_ISSUER_DOMAIN` — your Frontend API URL (e.g. `https://your-app.clerk.accounts.dev`)
@@ -26,17 +26,17 @@ Frontend uses Convex React hooks (`useQuery`, `useMutation`) with `ConvexProvide
 
 Backend is Fastify + Mastra. Fastify serves the HTTP API (Clerk JWT auth on protected routes via `backend/src/clerk-auth.ts`). Mastra (`backend/src/mastra/`) is the workflow orchestration layer — it wraps pipelines into inspectable workflows with a Studio UI. Both run as separate Docker services sharing the same source code.
 
-The schema inference pipeline: frontend calls `POST /infer-schema` → Fastify verifies the Clerk JWT → calls `inferSchema()` in `backend/src/pipeline/schema-inference.ts` → Claude Sonnet 4.6 via OpenRouter → returns a Zod-validated `DatasetSchema` → frontend maps it to editable columns in the wizard.
+The schema inference pipeline: frontend calls `POST /infer-schema` → Fastify verifies the Clerk JWT → calls `inferSchema()` in `backend/src/pipeline/schema-inference.ts` → selected LLM provider/model → returns a Zod-validated `DatasetSchema` → frontend maps it to editable columns in the wizard.
 
-The populate pipeline: frontend calls `POST /populate` with `{ datasetId, datasetName, description, columns }` → Fastify verifies the Clerk JWT → triggers `populateWorkflow` which: (1) clears existing rows, (2) builds a prompt from the schema, (3) runs the populate agent (Claude Sonnet 4.6) which searches the web via TinyFish APIs, then inserts rows into Convex one by one. Rows appear in realtime on the frontend via Convex reactive queries.
+The populate pipeline: frontend calls `POST /populate` with `{ datasetId, datasetName, description, columns }` → Fastify verifies the Clerk JWT → triggers `populateWorkflow` which: (1) clears existing rows, (2) builds a prompt from the schema, (3) runs the populate agent using the selected LLM provider/model. The agent searches the web via TinyFish APIs, then inserts rows into Convex one by one. Rows appear in realtime on the frontend via Convex reactive queries.
 
 Convex functions use `ctx.auth.getUserIdentity()` to get the authenticated user. The `ownerId` field on datasets stores `identity.subject` (Clerk user ID). Do not pass `ownerId` from the client.
 
 ## Environment Variables
 
 Root `.env` is the only local env file. Docker Compose, package scripts, and Convex CLI helper targets all read it. Key variables:
 - `TINYFISH_API_KEY` — used by the populate agent for web search and fetch (get one at https://agent.tinyfish.ai/api-keys?utm_source=github&utm_medium=organic&utm_campaign=bigset-developer-2026q2)
-- `OPENROUTER_API_KEY` — used by backend and Mastra for AI model calls
+- `OPENROUTER_API_KEY` — production default LLM provider key; local mode can use OpenRouter, OpenAI, Anthropic, or custom OpenAI-compatible via setup UI
 - `NEXT_PUBLIC_CLERK_PUBLISHABLE_KEY`, `CLERK_SECRET_KEY` — shared by frontend and backend
 - `CONVEX_SELF_HOSTED_ADMIN_KEY` — used by backend for system-level Convex writes
 

README.md

@@ -93,7 +93,7 @@ On first launch, BigSet sends you to setup. You'll connect two services:
 | Service | What it's for | Get your key |
 |---------|--------------|-------------|
 | **TinyFish** | Web search + page fetching | [tinyfish.ai/api-keys](https://agent.tinyfish.ai/api-keys?utm_source=github&utm_medium=organic&utm_campaign=bigset-developer-2026q2) |
-| **OpenRouter** | LLM calls (schema inference + agents) | [openrouter.ai/settings/keys](https://openrouter.ai/settings/keys) |
+| **LLM provider** | Schema inference + agents | OpenRouter, direct providers, Ollama, LM Studio, or another OpenAI-compatible endpoint |
 
 Local API keys are stored in your OS keychain.
 
@@ -214,23 +214,24 @@ Once everything is ready, you'll see:
 | **Mastra Studio** (workflow inspector) | [localhost:4111](http://localhost:4111) |
 
 Open [localhost:3500](http://localhost:3500). The setup screen will ask for
-TinyFish and OpenRouter credentials and save them to your OS keychain for this
-workspace.
+TinyFish credentials plus an LLM provider. Choose OpenRouter, a direct hosted
+provider, Ollama, LM Studio, or another OpenAI-compatible endpoint. BigSet saves
+local keys to your OS keychain for this workspace.
 
-### Step 3: Connect TinyFish and OpenRouter
+### Step 3: Connect TinyFish and an LLM provider
 
-TinyFish powers web search and page fetching. OpenRouter routes LLM calls to
-the models BigSet uses for schema inference and agents.
+TinyFish powers web search and page fetching. Your LLM provider powers schema
+inference and dataset-building agents.
 
 1. Create a TinyFish key at [agent.tinyfish.ai/api-keys](https://agent.tinyfish.ai/api-keys?utm_source=github&utm_medium=organic&utm_campaign=bigset-developer-2026q2)
-2. Create an OpenRouter key at [openrouter.ai/settings/keys](https://openrouter.ai/settings/keys)
-3. Paste both into BigSet's setup screen
-
-OpenRouter is pay-as-you-go; $5-10 is plenty to start.
+2. Choose an LLM provider. Hosted providers use an API key; Ollama and LM Studio
+   need no key by default.
+3. Save the provider, then choose models for each role in the model picker when
+   the provider does not supply defaults.
 
 > **Note:** root `.env` is the only local env file. If you edit Convex functions in `frontend/convex/`, run `make convex-push` to deploy the changes.
-
-> **Free tier:** cloud signed-in accounts get **2,500 row operations per calendar month** (resets on the 1st, UTC). Local mode bypasses the cloud quota and uses your TinyFish/OpenRouter accounts directly.
+>
+> **Free tier:** cloud signed-in accounts get **2,500 row operations per calendar month** (resets on the 1st, UTC). Local mode bypasses the cloud quota and uses your TinyFish + LLM provider accounts directly.
 
 ### Step 4 (optional): Load curated datasets
 
@@ -311,7 +312,7 @@ If you want a completely fresh start: `make clean` then `make dev`.
 | Auth | Local auth (dev); [Clerk](https://clerk.com) (cloud) |
 | Database | [Convex](https://convex.dev) (self-hosted) |
 | Data Collection | [TinyFish](https://www.tinyfish.ai?utm_source=github&utm_medium=organic&utm_campaign=bigset-developer-2026q2) APIs (Search, Fetch, Browser) |
-| AI orchestration | [Mastra](https://mastra.ai) workflows + [Vercel AI SDK](https://sdk.vercel.ai) + [OpenRouter](https://openrouter.ai) → Claude Sonnet (schema inference + populate agent) |
+| AI orchestration | [Mastra](https://mastra.ai) workflows + [Vercel AI SDK](https://sdk.vercel.ai) + a configured hosted, local, or OpenAI-compatible LLM provider |
 | Table view | [TanStack Table](https://tanstack.com/table) + [react-window](https://github.com/bvaughn/react-window) virtualization |
 | Exports | CSV (built-in) + XLSX ([SheetJS](https://sheetjs.com), dynamic-imported) |
 | Analytics | [PostHog](https://posthog.com) — events, session replay, error tracking (optional) |

backend/CLAUDE.md

@@ -15,7 +15,7 @@ To add a new protected route, register it inside the scoped plugin in `src/index
 
 ## Schema Inference Pipeline
 
-`src/pipeline/schema-inference.ts` — takes a natural language prompt and returns a structured `DatasetSchema` (Zod-validated, defined in `src/pipeline/types.ts`). Uses Claude Sonnet 4.6 via OpenRouter (`@openrouter/ai-sdk-provider` + Vercel AI SDK). Auto-retries once on validation failure by feeding the error back to the model.
+`src/pipeline/schema-inference.ts` — takes a natural language prompt and returns a structured `DatasetSchema` (Zod-validated, defined in `src/pipeline/types.ts`). Uses the configured LLM provider/model via `src/config/llm.ts` + Vercel AI SDK. Auto-retries once on validation failure by feeding the error back to the model.
 
 The pipeline is a pure function (`inferSchema(prompt) → DatasetSchema`). It is called by both Fastify (for the HTTP API) and Mastra (for workflow orchestration).
 
@@ -26,7 +26,7 @@ The pipeline is a pure function (`inferSchema(prompt) → DatasetSchema`). It is
 - `src/mastra/index.ts` — registers workflows with the `Mastra` instance (the populate agent is built per-run, not registered as a singleton)
 - `src/mastra/workflows/infer-schema.ts` — `inferSchemaWorkflow`, a single-step workflow wrapping `inferSchema()`
 - `src/mastra/workflows/populate.ts` — `populateWorkflow`, 3-step workflow: clear rows → build prompt → run populate agent
-- `src/mastra/agents/populate.ts` — `buildPopulateAgent(authorizedDatasetId, authContext, columns)`, builds the orchestrator agent (Claude Sonnet 4.6) with 3 tools: `search_web`, `fetch_page`, `investigate_row`. No write access — all inserts go through investigate subagents.
+- `src/mastra/agents/populate.ts` — `buildPopulateAgent(authorizedDatasetId, authContext, columns)`, builds the orchestrator agent with the configured LLM provider/model and 3 tools: `search_web`, `fetch_page`, `investigate_row`. No write access — all inserts go through investigate subagents.
 - `src/mastra/agents/investigate.ts` — `buildInvestigateAgent(authorizedDatasetId, authContext, columns)`, builds a per-entity subagent with `insert_row`, `list_rows`, `search_web`, `fetch_page`. Researches one entity, inserts at most one row, returns structured feedback (`INSERTED/SUMMARY/CLUES/REASON`).
 - `src/mastra/tools/investigate-tool.ts` — `buildInvestigateTool(authorizedDatasetId, authContext, columns)` creates the `investigate_row` tool. The orchestrator calls it to hand off a lead; it spawns a fresh investigate agent, runs it (maxSteps: 25), parses the structured output, and returns it to the orchestrator. Errors are caught and returned as structured failures so the orchestrator can self-correct.
 - `src/mastra/tools/dataset-tools.ts` — `buildPopulateTools(authorizedDatasetId, authContext)` factory returning 5 Convex-backed tools: `insert_row`, `list_rows`, `get_row`, `update_row`, `delete_row`. The dataset id is captured by closure so the LLM cannot redirect writes to other datasets; `authContext` (Clerk userId + workflow run id) is captured for caller-attribution in security logs and the `CAPABILITY_VIOLATION` PostHog event. See the security note at the top of the file.
@@ -48,7 +48,7 @@ Required env vars (see `.env.example`):
 - `CONVEX_URL` — Convex instance URL
 - `CONVEX_SELF_HOSTED_ADMIN_KEY` — for system-level Convex writes (internal mutations)
 - `CLERK_SECRET_KEY`, `CLERK_PUBLISHABLE_KEY` — for JWT verification
-- `OPENROUTER_API_KEY` — for AI model calls
+- `OPENROUTER_API_KEY` — production default LLM provider key; local mode can use providers including OpenRouter, direct hosted APIs, Ollama, LM Studio, and custom OpenAI-compatible endpoints via the setup UI
 - `TINYFISH_API_KEY` — for web search and fetch (populate agent). Get one at https://agent.tinyfish.ai/api-keys?utm_source=github&utm_medium=organic&utm_campaign=bigset-developer-2026q2
 
 In Docker, these are interpolated from the root `.env` file via `docker-compose.dev.yml`.