MCP connector

Connect Atlas to your AI assistant

One stable address gives ChatGPT, Claude, Claude Code, or Copilot access to your organization's governed definitions of record — the meaning your team already agreed on. It reads them; the one thing it can write is a conversation handing a question back to your team. It never changes a definition, never touches warehouse data, and only ever sees what you can.

Stable public URL Read-only, one exception Deep research ready OAuth 2.1 · PKCE

No subdomain to look up — you pick your organization while you connect.

https://datacoves.ai/mcp

Governed meaning, not another integration

Atlas publishes your Knowledge Base as an MCP server your assistant reads from. It pulls the agreed definition when it needs it — Atlas enforces nothing on your tools — and when a definition looks wrong, it can hand the question back to your team instead of guessing.

Pull, not control

An analyst may ground on the governed definition, or ignore the connector entirely. Atlas makes the single source of truth available and access-controlled — it governs meaning, never action.

Scoped to your reach

Every read is checked against your own permissions. A page you cannot read simply looks not-found — with no hint that it exists.

Bound to one org

The token you approve is bound to the organization you picked, so it can never reach another organization's knowledge — on any address.

What happens when you connect

Four steps, all at the Atlas front door.

  1. 1

    Add the URL

    Your client registers itself

    Paste the Atlas address into your assistant. It discovers the Atlas authorization server and registers itself automatically (dynamic client registration) — there is no client id or secret for you to copy anywhere.

  2. 2

    Prove your email

    Sign in at the Atlas front door

    Atlas asks you to prove control of your work email — your company's single sign-on, or a one-time code. Passwordless, always: Atlas stores no passwords.

  3. 3

    Pick your org

    Choose which organization to connect

    If your email reaches more than one Atlas organization, you pick one here. This is why there is no subdomain to look up — the address is the same for everybody, and the choice happens at consent time.

  4. 4

    Approve read access

    Approve read-only knowledge access

    You see exactly one scope — knowledge:read — and approve it. The token is bound to the organization you picked, refreshes silently, and you can revoke it from Atlas at any time.

Set up your client

Same address everywhere. Pick your assistant.

Custom connector · Business, Enterprise, or Edu workspace

  1. 1 Open Settings → Connectors → Create custom connector.
  2. 2 Paste the Atlas address as the MCP server URL and choose OAuth authentication.
  3. 3 Save, then press Connect — Atlas walks you through the four steps above.
  4. 4 Atlas now appears as a source you can pick per conversation, or in deep research.

MCP server URL

{@active_client.code}

Make Atlas the source of truth

The connector is pull, not control — it cannot force a client to ground on it. Paste this rule into your client's own instructions so it always checks Atlas first.

Atlas is our organization's governed source of truth for data definitions, metrics, and reference docs. Before answering any data or analytics question, search Atlas (search_knowledge / get_page) and ground your answer in what you find. Prefer Atlas definitions over assumptions, and say so when Atlas has no answer.

What it exposes

Every tool, what it takes, and exactly what it returns. All read-only but two: start_definition_conversation hands a page back to your team by opening an Atlas conversation about it, and start_conversation asks for a definition that does not exist yet — neither changes a definition, and they share one extra approval. Two others — search and fetch — are ChatGPT's deep-research compatibility pair: thin aliases over the same read API, so the tools they mirror keep their names and shapes.

  • list_workspaces Read-only

    List the workspaces you can reach.

    Input
    No arguments.
    Returns
    Every workspace you can reach, each as id · name. A workspace outside your reach never appears — it is the starting point for narrowing any later call.
  • list_areas Read-only

    List the areas inside a workspace, to narrow a search.

    Input
    workspace (required) — an id or name from list_workspaces.
    Returns
    The workspace's areas, each as id · name — the ids you pass as the optional area ceiling on a search.
  • get_page_tree Read-only

    Show a workspace's knowledge base as an indented map of its pages.

    Input
    workspace (required) — an id or name from list_workspaces.
    Returns
    An indented map of the workspace's knowledge base by hierarchy. A page you may not read appears as a title-only node.
  • search_knowledge Read-only

    Semantic search across governed definitions, reference docs, and notes.

    Input
    query and workspace (both required); optional area ceiling, type, and limit.
    Returns
    The pages you may read, most relevant first, each citing its title, type and current version. Reach- and confidentiality-filtered before it ranks. A hit whose concept has a documented variant — another definition deliberately kept different — is flagged, with the reason on the page read. A hit that is the definition of record for a physical table also names that table, fully qualified as catalog.schema.table.
  • get_page Read-only

    Fetch the full body of a single page by id or slug.

    Input
    id, or slug plus workspace.
    Returns
    The page's current version — its full markdown body, cited by title, type and version, followed by a note naming any documented variant of the concept and the reason for the split, and a note naming the physical table(s) the page is the definition of record for. A page that is missing and one you may not read return the identical not-found result.
  • get_history Read-only

    Show a page's version history, with a diff of the two most recent versions.

    Input
    page (required) — the page id.
    Returns
    Each version's number and timestamp, plus a line diff of the two most recent — how a definition of record evolved.
  • find_definition_for_table Read-only

    Ask which definition of record governs a physical warehouse table.

    Input
    table and workspace (both required) — the table fully qualified as catalog.schema.table; optional area ceiling.
    Returns
    The definition(s) of record that govern that table and you may read, each citing its title, type and current version. You never name a connection: Atlas answers for every warehouse connection you can reach, so more than one definition can come back when the same table name exists in more than one warehouse — and no connection is ever named, counted or hinted at. A table nothing governs and one whose definition you may not read return the identical result.
  • fetch Read-only

    ChatGPT deep-research compatibility: read one page by the id a search returned.

    Input
    id (required) — a page id from a search result.
    Returns
    That page's full text, with its title and type, and — where known — an absolute URL and its current version. Its text carries the same documented-variant and governed-table notes get_page appends. A page that is missing and one you may not read return the identical error.
  • start_definition_conversation Starts a conversation

    Hand a page back to its owners: start an Atlas conversation about it and post the first message.

    Input
    page and message (both required); optional origin_url and origin_label naming where the request came from.
    Returns
    A link to the new conversation, and its id. The page is not changed, no change request is created, and Atlas does not reply — a person opens the room and answers. It writes, so it needs conversation access approved on the connector.
  • start_conversation Starts a conversation

    Ask for a definition that does not exist yet: start an Atlas conversation in a workspace and post the first message.

    Input
    workspace and message (both required); optional area to file it under, and optional origin_url and origin_label naming where the request came from.
    Returns
    A link to the new conversation, its id, and the area it was filed under. No knowledge page and no change request are created, and Atlas does not reply — a person opens the room and answers. Like start_definition_conversation it writes, so it needs conversation access approved on the connector.

Want the longer reference? Read the MCP documentation

Security & data flow

The factual disclosure, in plain terms — the same one our Privacy Policy restates.

What crosses the wire

  • Your question, as the assistant's search text.
  • Governed Knowledge Base content you can already read — definitions of record, reference docs, notes.
  • Page titles, ids, areas, version history, and the links between pages.
  • An OAuth access token bound to the one organization you approved.

What never leaves Atlas

  • Warehouse data. The outlet cannot run a query or reach a table — ever.
  • Secrets. Passwords, keys, and connection configuration never leave Atlas.
  • Anything outside your reach. A page you cannot read looks not-found, with no hint it exists.
  • Changes to your knowledge. No edit, promote, or delete — the only thing the outlet can create is a conversation asking your team about a page, and only if you approve it.
  • Another organization's knowledge. The token's binding makes it unreachable.

Connector facts

The listing details a directory reviewer asks for.

Listing name
Atlas
Publisher
Datacoves
One-liner
Search and read your organization's governed data definitions, and hand a question back to the team that owns one.
Authentication
OAuth 2.1 + PKCE, dynamic client registration, refresh and revocation.
Scope requested
knowledge:read
Support
support@datacoves.com, or the contact form.
Availability
Worldwide, except where prohibited by applicable export-control or sanctions law.
Data retention
No assistant conversation content is stored by the connector.

Don't have Atlas yet?

Spin up a workspace, bring your data and knowledge into one canvas, then connect it to the assistant you already use.

Privacy Policy Terms of Service