{"slug": "wren-ai-is-now-available-in-the-claude-directory", "title": "Wren AI Is Now Available in the Claude Directory", "summary": "Wren AI is now listed as a connector in the Claude Directory, allowing Claude to query governed data through Wren AI's context layer with OAuth authorization and row-level security. The connector exposes eight tools, including ask, generate_sql, and respond_clarification, and integrates with Wren AI's MCP server v0.53, reducing setup time to about a minute.", "body_md": "[The Wren Journal](/blog)\n\n# Wren AI Is Now Available in the Claude Directory\n\nWren AI is now listed in the Claude Directory. Add the connector inside Claude, authorize with OAuth, and Claude asks questions of your context layer instead of your raw schema, with row-level security and a full audit trail at query time.\n\nWren AI Product Team\n\nUpdated: Aug 09, 2026\n\nPublished: Aug 09, 2026\n\nWren AI is now listed as a connector in the [Claude Directory](https://claude.ai/directory/connectors/wren-ai). You add it from inside Claude, authorize it with OAuth, and Claude can ask questions of your governed data. No JSON config file, no warehouse credential pasted into a client, no separate deployment step.\n\nWe shipped the Wren AI MCP server in v0.53. Everything since then has gone into the part users actually feel: how long it takes to get from \"we should try this\" to a governed answer in the chat window. The directory listing closes that gap to about a minute.\n\nClaude can already reach your data through MCP. What it reaches through determines whether the answer is governed. The connector puts your context layer in that position by default.\n\n## What the connector exposes\n\nThe connector does not hand Claude a set of tables. It exposes your context layer: the modeled business definitions, the sanctioned join paths, and the access rules, enforced when the query runs.\n\nThree things change compared with pointing an agent at a database directly.\n\n**1. Meaning is resolved, not inferred.** Business definitions live in MDL, the Wren AI context model, stored as git-versioned files in your own repository. When a definition changes in a pull request, every agent picks it up on its next question. Claude never has to work out what `rev_net_adj`\n\nmeans from the column name.\n\n**2. Joins are declared, not discovered.** Relationships are part of the model. Our Rust engine, built on Apache DataFusion, plans SQL from those declared paths across 20+ sources including Postgres, BigQuery, Snowflake, and Databricks. The same question resolves to the same join for every caller.\n\n**3. Every request carries an identity.** Claude authenticates through OAuth rather than a shared service account. Row and column level security applies at query time to the authenticated caller, and each request is traceable to who asked, what it resolved to, and the SQL that ran. Access is scoped and revocable.\n\n## The eight tools\n\nThe connector ships eight tools: `ask`\n\n, `generate_sql`\n\n, `run_sql`\n\n, `generate_chart`\n\n, `generate_summary`\n\n, `get_project_metadata`\n\n, `list_projects`\n\n, and `respond_clarification`\n\n.\n\n`respond_clarification`\n\nis the one worth calling out. It lets the connector ask a question back. When a request maps to two different metrics, the useful behavior is not to pick one and commit to it, but to say which two and let the user choose. A raw SQL tool has no way to express that, because it has no model of what the alternatives are. A context layer does.\n\n`get_project_metadata`\n\nand `list_projects`\n\nserve a related purpose. Claude can see which models, metrics, and relationships exist before it composes anything, so its first move is to read the available vocabulary rather than guess at it.\n\nHere is the flow end to end, from adding the connector to a governed answer coming back.\n\n## Setting it up\n\nEnable the MCP connection on your Wren AI project, then add Wren AI from the Connectors section of the Claude Directory and authorize it.\n\nAuthorization runs through an OAuth proxy. A database credential never reaches the client, and you can revoke a client's access without rotating anything in your warehouse. Claude holds a token scoped to a project, not a key to the data source.\n\nFor teams already running the MCP server from a config file, the connector is the same server reached a shorter way. Existing projects, models, and policies carry over.\n\n## One boundary for every client\n\nMost teams we work with are not deploying a single agent. They have Claude for the analysts, an IDE assistant for the engineers, and a copilot inside their own product, with more arriving each quarter.\n\nWhen each client connects to the database on its own, each one carries its own interpretation of the business, its own join choices, and its own access configuration. Definitions drift between them, and reconciling the results becomes recurring work.\n\nRouting them through one MCP boundary changes the arithmetic. The context layer is defined once. Policy is enforced once, at execution, for every caller. Wren AI runs the same boundary for people and agents, so a question asked in the Wren AI UI, through the API, or from Claude over MCP resolves the same way.\n\n| Direct database connection | Wren AI connector | |\n|---|---|---|\n| Business definitions | inferred per client | resolved from MDL, versioned in git |\n| Joins | chosen by the model | declared in the context layer |\n| Access control | scope of the service account | RLS and CLS at query time, per caller |\n| Audit trail | one shared query log | every request traced to an identity |\n| Adding a second agent | repeat the whole setup | authorize another client |\n\n## Availability\n\nThe connector is live in the Claude Directory now, listed under Data & Analytics and Productivity. It is published in the community tier, which means it has passed Anthropic's automated review rather than their verified-partner review.\n\nWren AI core is [open source](https://github.com/Canner/WrenAI). The MCP server, the MDL context contract, and the engine are all inspectable, and you can run the whole path locally against your own Postgres before connecting anything in production.\n\nIf you are the person who has to sign off on this, the short version for the review ticket: Claude never touches the database, it calls an MCP server that authenticates the client over OAuth with scoped and revocable access, definitions come from a git-versioned context model, row and column security is enforced at query time per caller, and every request is logged with its caller, its resolved plan, and the SQL that executed.\n\nAdd the connector from the [Claude Directory](https://claude.ai/directory/connectors/wren-ai), or read the [technical background on the Wren AI MCP server](https://www.getwren.ai/post/wren-ai-mcp-server-governed-agent-access).\n\n### Supercharge your data with AI today\n\nJoin thousands of data teams already using Wren AI to make data-driven decisions faster and more efficiently.\n\n[Start Free Trial](https://cloud.getwren.ai/)", "url": "https://wpnews.pro/news/wren-ai-is-now-available-in-the-claude-directory", "canonical_source": "https://getwren.ai/post/wren-ai-claude-directory-connector", "published_at": "2026-08-09 00:00:00+00:00", "updated_at": "2026-08-09 12:54:52.049238+00:00", "lang": "en", "topics": ["artificial-intelligence", "ai-products", "ai-tools", "ai-infrastructure"], "entities": ["Wren AI", "Claude Directory", "Claude", "Apache DataFusion", "Postgres", "BigQuery", "Snowflake", "Databricks"], "alternates": {"html": "https://wpnews.pro/news/wren-ai-is-now-available-in-the-claude-directory", "markdown": "https://wpnews.pro/news/wren-ai-is-now-available-in-the-claude-directory.md", "text": "https://wpnews.pro/news/wren-ai-is-now-available-in-the-claude-directory.txt", "jsonld": "https://wpnews.pro/news/wren-ai-is-now-available-in-the-claude-directory.jsonld"}}