Wren AI MCP

Agentic Mode not supported

Wren AI MCP currently runs in Interactive Mode only. Agentic Mode is not supported. Calls through MCP will use Interactive Mode even if the connected project has Agentic Mode available.

Using Claude or ChatGPT? Install from the app directory instead

Wren AI is published in Claude's connector directory (claude.ai/customize/connectors/id/wren-ai), so on Claude web/desktop you can connect in one click without pasting the endpoint URL. Wren AI is published in the ChatGPT plugin directory too. See the Claude.ai Integration and ChatGPT Integration guides. Adding the endpoint as a custom connector still works, and is what you want for self-hosted or non-default endpoints.

Overview

Wren AI MCP connects your AI agent (Claude, ChatGPT) to Wren AI so you can query your data, generate SQL, build charts, and get business insights in natural language — right from the tools you already use.

It's a single, organization-level endpoint secured with OAuth:

• Endpoint: https://cloud.getwren.ai/api/mcp
• Auth: OAuth (your Wren AI account) — no API key and no local proxy
• Plan: Essential Cloud and above

How It Works

1. An organization admin enables MCP for the organization and chooses which projects are exposed.
2. You add the one endpoint URL to your AI client and log in with OAuth.
3. Your agent calls list_projects to discover the projects you can reach, then passes a projectId to the other tools.
4. If your organization uses row/column-level security (an Enterprise Cloud feature), it's applied per user, server-side — you only ever see what your Wren AI account is allowed to see.

Before You Begin

1. Wren AI Essential Cloud plan and above.

2. An org admin enables MCP and exposes projects. In Wren AI, go to Settings → Organization → MCP, turn on Enable MCP for this organization, and toggle Exposed for each project you want reachable. Optionally set New projects default to Exposed. The endpoint URL shown here — https://cloud.getwren.ai/api/mcp — is shared by everyone in the org.

   

   note

   Only exposed projects are reachable, and each user still only sees projects they're a member of. Turning a project off here (or disabling the org switch) immediately cuts its MCP access.

3. Claude connector permission (Team/Enterprise workspaces). On Claude's Team/Enterprise plan, an Owner must enable the Wren AI connector for the organization under Organization settings → Connectors before members can connect it. For a self-hosted endpoint added as a custom connector, your workspace admin must also enable the Add custom connectors permission for your account.

Claude (Web & Desktop)

Wren AI is published in Claude's connector directory, so you don't have to add the endpoint by hand. Auth is handled via OAuth — no API key needed.

Step 1 — Connect the Wren AI connector:

Go to claude.ai/customize/connectors/id/wren-ai and click Connect to Claude, then log in with your Wren AI account. Full walkthrough with screenshots: Claude.ai Integration.

tip

You can also find it from inside Claude: Customize → Connectors → Add → Browse connectors, then search WrenAI (or Wren).

Self-hosted or non-default endpoints

The listing points at Wren AI Cloud (https://cloud.getwren.ai/api/mcp). For a self-hosted Wren AI, add it as a custom connector instead: Customize → Connectors → Add custom connector, name it WrenAI, enter your endpoint URL, and Connect — then log in and authorize the app.

Step 2 — Upload the analyst skill (optional):

1. Open Customize → Skills → Create a skill → Upload a skill.
2. Upload web/claude/wrenai-analyst/SKILL.md.

tip

Generate the skill files locally:

git clone https://github.com/canner/wrenai-mcp.git && cd wrenai-mcp && ./setup --host web

ChatGPT

Wren AI is published in the ChatGPT plugin directory, so you don't have to add the endpoint by hand. Auth is handled via OAuth.

Step 1 - Install the Wren AI plugin:

Open the Wren AI plugin listing in ChatGPT and click Install plugin, then log in with your Wren AI account. Full walkthrough with screenshots: ChatGPT Integration.

tip

You can also find it from inside ChatGPT: open Plugins and search wren.

Self-hosted or non-default endpoints

The listing points at Wren AI Cloud (https://cloud.getwren.ai/api/mcp). For a self-hosted Wren AI, add it as a custom connector instead: Plugins → Connectors → New Plugin, keep Connection: Server URL, enter your endpoint, set Authentication: OAuth, tick "I understand and want to continue", then Create.

Step 2 - Add custom instructions (optional):

1. Go to Settings → Personalization → Custom Instructions.
2. Paste the content from web/chatgpt/wrenai-analyst-skill.md.

tip

Generate the skill content locally:

git clone https://github.com/canner/wrenai-mcp.git && cd wrenai-mcp && ./setup --host chatgpt

Claude Code (CLI)

claude mcp add --transport http wrenai https://cloud.getwren.ai/api/mcp

Then run /mcp inside Claude Code, select wrenai, and complete the OAuth login in your browser. No API key or proxy needed.

For the analyst skill and more, see the WrenAI-mcp repository.

Using It

Once connected, just ask in natural language. Behind the scenes the agent calls list_projects to choose a project, reads schema with get_project_metadata, then answers with ask / generate_sql / run_sql and renders a generate_chart visualization.

Example prompts:

• "List my Wren AI projects."
• "In the Sales project, what were the top 5 products by revenue last quarter? Chart it."

note

Each connection is bound to a single organization. list_projects shows only the exposed projects in that organization, even if your account belongs to other MCP-enabled organizations.

Rating an answer

info

Answer feedback is currently available on self-hosted Wren AI. Coming soon to Wren AI Cloud.

When an answer is wrong, or exactly right, say so in your own words, for example "That total is wrong, it should exclude refunds" or "Perfect, that's the number I needed." The agent records your verdict on the Wren AI answer you're talking about with submit_feedback, so the people who review answer quality can see which answers missed and why. The agent doesn't ask you to confirm first, although your client may ask you to allow the tool.

• One report per verdict. A report can't be withdrawn or edited. If you change your mind, say so: that is recorded as a new report and the earlier one stays. Sending an identical report again doesn't create a duplicate.
• What can be rated. Answers from ask, generate_sql, generate_summary and generate_chart, including a call that failed when its error includes an answerId.
• When a report isn't recorded, the agent tells you, and offers to retry a temporary failure. An answer given before rating was available, or one that failed before Wren AI started on the question, can't be rated; ask the question again and rate the new answer.

Tools

Tool	Purpose	Takes projectId
list_projects	Projects you can query	—
get_project_metadata	Table schema for a project	✓
ask	End-to-end natural-language Q&A	✓
generate_sql	Question → SQL	✓
run_sql	Execute SQL	✓
generate_chart	Vega chart spec	✓
generate_summary	Summarize results	✓
respond_clarification	Answer a follow-up question from Wren AI	✓
submit_feedback	Record whether an answer was right or wrong (coming soon to Wren AI Cloud)	✓

ask, generate_sql, generate_summary and generate_chart return an answerId with each answer. The agent passes it to submit_feedback so the report is attached to that exact answer; you never need to copy it.

Data Security

Row/column-level security (RLS/CLS), an Enterprise Cloud feature, is enforced per user: policies are resolved server-side from the Wren AI account that authenticated the MCP connection (the user who logged in via OAuth), and the connected client cannot override them. Each person sees only the rows and columns their own account is permitted to see.

Watch out for shared accounts

Because data access is resolved from the connecting account, anyone using a shared Wren AI account inherits exactly that account's data permissions — so RLS/CLS won't distinguish between the individual people behind it. Have each person connect with their own Wren AI account so security policies apply to the correct identity.

Answer ratings. When you rate an answer, the report carries your own words, including any corrected value, along with the account that connected. The agent sends it without a confirmation step, so leave out anything you wouldn't want the people who review answer quality to see. Each report also appears in the project's API History.

Self-hosted

On a self-hosted Wren AI, answer ratings aren't sent to Wren AI. They're kept in the deployment's log archive and leave only inside an export from Settings → Project → Export logs (project owners) that you choose to share, or go to your own Langfuse if you've connected one. If no feedback storage is configured, the agent tells you the rating couldn't be recorded.

Migrating from the per-project URL

The legacy per-project URL — https://cloud.getwren.ai/api/mcp/{token} — and its Settings → Project → MCP connection page are deprecated and will be removed soon. To migrate:

1. Have an admin enable MCP and expose the project in Settings → Organization → MCP (see Before You Begin).
2. In your AI client, replace the old per-project URL with https://cloud.getwren.ai/api/mcp and reconnect.
3. Your agent now selects the project via list_projects instead of the URL.