MCP Server

The Querri MCP server lets an AI assistant work with your Querri data on your behalf. Once it’s connected, you can ask about your data, start an analysis project, and look at sources, views and dashboards without leaving the assistant.

What is MCP?

MCP (Model Context Protocol) is an open standard for connecting AI assistants to outside systems. You add a server to your assistant, sign in once, and the assistant gets a set of tools it can call while you talk to it.

Connect your assistant

The server lives at:

https://app.querri.com/mcp

It speaks MCP over streamable HTTP and signs you in with OAuth (authorization code with PKCE). When you connect, the assistant opens Querri’s normal sign-in page, and once you approve, it acts as you.

Sign-in only returns to callback addresses Querri allows:

• Claude’s MCP callbacks on claude.ai and claude.com
• local callbacks on localhost or 127.0.0.1, which desktop and command-line clients use

A hosted assistant with any other callback can’t finish signing in.

Claude

1. In Claude, open Settings and go to Connectors.
2. Add a custom connector.
3. Enter https://app.querri.com/mcp as the server URL.
4. Sign in to Querri when prompted.
5. Start a new conversation. Querri’s tools show up when they’re relevant.

Other MCP clients

A client that accepts a remote MCP server URL and signs in through a local callback connects the same way. Add the URL as a remote server and choose OAuth as the authentication method.

What the assistant can see

The assistant acts as the person who signed in. It sees what you can see in Querri, and access policies filter the rows it reads, just as they do for you.

If you belong to more than one organization, the connection uses the one you signed in to. list_organizations shows every organization on your account and which one is active. To use a different one, disconnect Querri in your assistant, connect again, and sign in to the other organization.

Available tools

Asking questions

Tool	What it does
ask	The default for almost every question. It runs the Librarian, which searches your Library and queries existing views and sources, then returns the answer and what it was based on in one call. Pass the returned conversation_id on a follow-up to keep the context.
ask_source	Answers a question about one source you already have the ID for, without using the Library.

Sometimes ask needs one thing settled before it can answer, such as which of two similar sources you meant. It then replies with a question and options instead of an answer. The assistant should pick from what it already knows, or ask you. And when a question needs transformation work the Library can’t do, ask says so, and the assistant moves on to a project.

Projects

Tool	What it does
start_project	Starts an analysis project for work like joining raw sources that have no view yet, building charts or reshaping data. It returns a project_id and chat_id straight away, and the analysis runs on the server.
get_project_answer	Reads the answer in a project chat. Each call waits up to about 25 seconds for a running answer, so the assistant calls it again while the status still says running.
list_projects	Lists analysis projects.
get_project	Returns a project as a diagram of its steps, plus a table of step IDs.
list_project_steps	Lists a project’s steps.
get_step_data	Returns the rows a step produced, and its chart if it has one.
create_project	Creates an empty project. start_project is usually what you want.

Sources, views, dashboards and files

Tool	What it does
list_sources	Lists data sources.
get_source	Returns a source’s columns and their types.
query_source	Runs SQL against one source, which is exposed as a view named data. Write SELECT ... FROM data, never the source’s display name. It can’t join across sources.
list_views	Lists SQL views.
get_view	Returns a view’s definition and details.
list_dashboards	Lists dashboards.
get_dashboard	Returns a dashboard’s details, including its widgets.
list_files	Lists uploaded files.

Writing to the Library

Tool	What it does
curate	Records knowledge in the Library: facts, collections, refining questions and confirmed KPIs.

curate only appears when your organization allows it. An admin turns on MCP clients may write to the Library, which is off by default. Without it, the connection is read-only and curate isn’t in the tool list. start_project and create_project stay available either way.

Organizations

Tool	What it does
list_organizations	Lists your organizations and shows which one is active.
switch_organization	Explains how to reconnect to a different organization.

Credits

Questions through ask use credits the same way a question to the Librarian in Querri does. If you’ve used all of your monthly credits, ask refuses the question instead of running it.

Tips for good results

• Ask the business question, not the steps. “Compare revenue by region this year and last” works better than describing joins and groupings.
• Keep one conversation going. Follow-ups that reuse the conversation_id keep what ask already worked out.
• Use SQL when you know the shape. If you know which source and what you want from it, query_source is the direct route.
• Don’t join by hand. Paging raw rows out of several sources to join them in the assistant is slow and easy to get wrong. Use ask, or start_project for real transformation work.

Troubleshooting

The assistant can’t connect. Reconnect Querri from the assistant’s connector settings and sign in again.

Sign-in fails on the redirect. The assistant’s callback address isn’t on the allowed list. See Connect your assistant.

The tools don’t appear. Some assistants only reach for a connector when the prompt is clearly about it. Try “Use Querri to…”.

curate is missing. Your organization hasn’t allowed MCP clients to write to the Library, so the connection is read-only.

A project answer seems stuck. Long analyses are normal. The assistant should keep calling get_project_answer until the status stops saying running.

It’s using the wrong organization. Disconnect, connect again, and sign in to the organization you want.

Building your own integration

If you’re writing Python that talks to Querri directly, use the Python SDK. To work from a terminal or a shell script, see the CLI.