Agent connections
Let Claude, Cursor or another AI assistant query the companies and datasets you choose.
What an MCP connection is
An MCP connection is how an AI assistant such as Claude connects to Gatlas. MCP, the Model Context Protocol, is the open standard AI assistants use to reach outside data, and Gatlas runs an MCP server. Each MCP connection is one key to that server. It decides what the assistant can read, which views it lists, and whether it may change anything, and every call made with it is logged.
Admins and above create connections, on the MCP page in the sidebar. That page also lists the connections you can see, when each was last used and, for connections made there, its setup link.
Create a connection
Start a new connection
Go to MCP and choose New connection.
Name it, and say who it is for
Give it a Connection name you will recognize later, such as the person and the assistant it is for. Add a Recipient name and Recipient email if Gatlas should email that person the setup instructions.
Choose the permissions
Under Permissions, choose Read-only or Read/write. New connections start as Read/write. See Read-only or read/write below.
Limit what it can see
Under Published views you can limit the connection to particular views. With one company, the connection reads your business and the form does not ask which company.
With several companies, choose under Companies what the connection can see:
- All companies: every company in the organization, including ones added later. If your own access is limited to some companies, this reads All your companies and means those.
- Companies and everything under them: each company you tick, plus every company under it, including ones added under it later.
- Only specific companies: exactly the companies you tick. A company added under one of them later is left out until you tick it too.
If your own access is to specific companies only, the middle choice is not offered, since it would reach companies you don't have.
Important
list_tables, describe_table and query_sql. To keep a company's data away from an assistant, leave that company out of the scope.Create, and copy what it shows
Choose Create. Gatlas shows the Server URL, the Key and a Setup link for recipient, which is valid for 30 days. You can find the key again later in the connection's Setup instructions on the MCP page, and the setup link stays there too, even after it expires. Anyone in the organization who can open the MCP page can read the key, so share access to Gatlas with that in mind.
Read-only or read/write
Every connection is either Read-only or Read/write, as chosen when it is created, and new connections start as Read/write. You can change it later with Edit access on the connection.
- Read-only reads data and sends feedback to Gatlas. It cannot change anything.
- Read/write can also make changes inside the connection's scope, for example: create, edit and delete views and artifacts, edit documents, run and schedule prompts (each run calls an AI model), reschedule syncs, and create setup links. Starting a sync needs sync management, which a connection made on the MCP page does not have. The tools that change something are marked below.
Choose Read-only for anyone who only asks questions. Neither kind can change anything in the tools your data syncs from: no MCP tool writes to them.
Tools take your business by default
With one company in a connection's scope, the assistant never has to name it: tools that act on a company, such as creating a view or a setup link, or renaming and describing your business, take that company when none is named. With several companies in scope, a setup link, a view from a template or a rename needs one named, and a call without one is refused rather than guessed; a view built from SQL with none named is saved untagged, readable only by connections for the whole organization.
Stop a connection
Open the connection on the MCP page and choose Delete connection. Its key stops working at once. A person who leaves the organization, or is removed from it, loses the connections they created there.
Every tool an assistant can use
An assistant sees these tools when it connects. On a Read-only connection, the tools marked Changes data are still listed, but the assistant is refused if it calls one. The sync tools that need a connection that can manage syncs are left out of any other connection, as the note under the list explains. Tools marked Changes data need a Read/write connection. You never call them yourself: you ask in plain language and the assistant picks the tool.
Getting started
get_agent_context- Loads the rules for working with your data, the companies in reach and any freshness warnings. Assistants call it first.
get_help- Explains how to use Gatlas: where to start, which views and documents to set up, and what to do when a question goes wrong.
list_entities- Lists the companies this connection can see.
submit_feedback- Sends feedback about Gatlas, such as a bug or missing data, to the Gatlas team.
Views and numbers
list_views- Lists every published view this connection can query.
search_views- Finds published views by keyword, company or topic.
describe_view- Reads a view's description, AI context and column notes before querying it.
query_view- Reads rows from a published view, with filters. The usual way to get numbers.
query_sql- Runs a read-only SQL query over the synced tables, for questions no view covers.
list_tables- Lists the raw synced tables for each company, with their columns.
describe_table- Describes one raw synced table: its columns, row count and date range.
create_view- Creates and publishes a view from a catalog template or from SQL, with its description and AI context.
edit_view- Changes a published view's name, SQL, AI context or column notes.
delete_view- Permanently deletes a view. Without a confirmation it only reports what would be affected.
Documents and files
list_documents- Lists the curated documents this connection can read, such as company overviews and board summaries.
get_document- Reads one curated document, its past versions and the source files behind it.
search_documents- Searches the raw synced files, such as contracts and board packs, when no curated document covers the question.
get_raw_document- Reads the text of one raw synced file, a page at a time.
edit_document- Updates a curated document and keeps the previous text as a version.
list_transform_prompts- Lists the prompts that write and refresh curated documents.
create_transform_prompt- Creates a prompt that writes a curated document from your files, documents and views, on a schedule or on demand.
edit_transform_prompt- Changes a prompt's instructions, inputs, output document or schedule.
run_transform_prompt- Runs a prompt now, as a preview or for real.
Sources and syncs
list_connections- Lists each company's sources with their sync schedule, status and how fresh the data is.
get_sync_status- Shows sync health for every source this connection can see. Needs a connection that can manage syncs.
get_recent_syncs- Lists the latest sync runs and how each one went. Needs a connection that can manage syncs.
trigger_sync- Starts a sync of one source now. Needs a connection that can manage syncs.
edit_sync_schedule- Changes how often a source syncs by itself.
create_entity- Adds a company to your organization, so its sources and views have a home. Only after you confirm its name.
edit_entity- Renames a company or rewrites its description and AI context. Only after you confirm the change.
mint_setup_link- Creates a setup link that someone opens with no Gatlas login, to connect one or more tools to a company or reconnect one. Only after you ask for one.
Artifacts
list_artifacts- Lists the published artifacts: dashboards, documents and decks built on your views.
list_components- Lists the saved charts, tables and text blocks artifacts can reuse, and how many artifacts use each one.
describe_artifact- Reads one artifact's charts, tables and text, the views they use and its share link.
describe_component- Reads one saved chart, table or text block and lists the artifacts that use it.
create_artifact- Builds a dashboard, an A4 document or a slide deck from charts, tables and text over published views.
create_component- Saves one chart, table or text block to the library, so any artifact can reuse it.
create_edition- Makes a numbered edition of a deck or document: a frozen copy at its own link, which an admin can withdraw.
edit_artifact- Changes an artifact's contents, layout or share link.
edit_component- Changes a saved chart, table or text block in every artifact that uses it.
delete_artifact- Permanently deletes an artifact, its share link and its editions. Without a confirmation it only reports what would be affected.
delete_component- Removes a saved chart, table or text block that no artifact uses anymore. Without a confirmation it only reports what would be affected.
The three sync tools, sync status, recent syncs and starting a sync, need a connection that can manage syncs, which the MCP page does not offer yet. To get one, write to team@gatlas.ai.