Tessera Studio

Tessera Studio is the local web workspace for exploring and analyzing a PostgreSQL, MySQL, SQLite, Turso, or MongoDB database. Start it with the executable published by @open-tessera/studio:

npx @open-tessera/studio@latest postgresql://readonly:[email protected]:5432/warehouse

What Studio does

• Chat with a database assistant using grounded text, tool state, and Evidence.
• Inspect the available catalog and run governed, read-only analysis.
• Show compact execution progress and query evidence alongside the response.
• Accept image attachments in a chat message, including pasted images.
• Keep local chat sessions in the Studio SQLite store.

The Studio server and its local session store run on Node.js 24 or later and Bun 1.3 or later. A project can keep the connection string and model-provider settings in a server-only tessera.config.ts file and then run npx @open-tessera/studio@latest from that project directory.

SQLite accepts an existing local database through file: or sqlite:. Turso accepts libsql: or turso: and reads its separate server-only credential from database.authToken or TURSO_AUTH_TOKEN. SQLite and Turso connections are always read-only in Studio.

npx @open-tessera/studio@latest file:/absolute/path/to/warehouse.db
TURSO_AUTH_TOKEN=... npx @open-tessera/studio@latest libsql://warehouse-org.turso.io

Vercel AI Gateway

In Settings → Model, choose Vercel AI Gateway, select a model, and enter a Gateway API key (or set AI_GATEWAY_API_KEY on the server). Leave Base URL empty for Mastra's native Vercel integration. The settings and chat model pickers load Vercel's public catalog of text models with tool support.

For a project configuration:

import { defineTesseraConfig } from "@open-tessera/studio";

export default defineTesseraConfig({
  database: { url: process.env.DATABASE_URL! },
  llm: { model: "vercel/openai/gpt-4.1-mini" },
});

OpenRouter remains the initial default. Switching gateways does not carry API keys or custom endpoints between providers. A custom Base URL uses the OpenAI-compatible chat completions protocol.

Jev evaluation

The Vercel model settings also include a Jev evaluation panel. It calls typesafe-ai/jev through AI SDK experimental_evaluate, independently of the chat model. Enter text or JSON state and a JSON object of boolean, choice, or score questions, then click Run Jev evaluation. Answers and token usage are shown inline. The evaluation always uses Vercel's native gateway endpoint and never saves or replaces the active chat configuration.

The same server-side AI_GATEWAY_API_KEY or Vercel key entered in the form is used for both modes. If needed, create a key with vercel ai-gateway api-keys create --name tessera, then configure it locally.

Local development

From this directory:

bun install
bun run dev

Nitro serves both the API and Vite client at http://127.0.0.1:4317. Run bun test src for the Studio test suite or bun run build to create the CLI, library, and deployable Nitro server.

The Nitro production output is written to dist/nitro. Start it with Node.js 24 or Bun:

node dist/nitro/server/index.mjs

The standalone studio CLI remains available and uses the same H3 application and Studio runtime as the Nitro deployment.

The package exports the server factory and the ./main command entry for integrations that embed Tessera Studio.

Continual harness

Studio uses an independent Reviewer/Planner harness to turn evidence-backed corrections into thread-local editable domain memory. Automatic review is enabled by default after 25 successful turns, with a 20-minute cooldown. It cannot edit prompts, tools, database permissions, approval policy, SQL boundaries, credentials, or connections.

studio: {
  continualHarness: {
    enabled: true,
    autoReviewInterval: 25,
    autoReviewCooldownMs: 20 * 60_000,
  },
}

Automatic changes remain isolated to the current thread. Cross-session memory requires an explicit Host promotion through the exported continual-harness API. See the architecture guide for the state machine, validation policy, rollback flow, and Sandbox boundary.