Base44
Build and edit Base44 apps from an AI assistant, and query data your app's users submit.
Query MongoDB collections and manage Atlas infrastructure.
Opens Agent Studio, where connecting is one click. The connector URL below works in any MCP client.
Paste it into any MCP client. Setup docs
The MongoDB Atlas MCP server lets Claude, ChatGPT or any MCP-compatible agent both query the documents in your MongoDB collections and manage the Atlas infrastructure holding them. Anthropic's directory lists 41 tools: full document CRUD on one side, and cluster, database-user and IP-access-list management on the other. Sign-in is OAuth against Atlas, and an Organization Owner must switch it on.
Verified connector
Listed by Anthropic as a partner connector in its Connectors Directory.
Connection checked by Agentman on .
Anthropic states this reflects the level of review a connector received, not a security audit.
Tool names from Anthropic's directory listing. This server requires sign-in, so we could not read tool descriptions or parameter schemas.
delete-many, drop-collection and drop-database are among the directory's 41 tools. This connector can destroy data, not merely read it.readOnlyHint and destructiveHint as advisory metadata sent to the client, which the client can ignore, and which authorize, restrict and block nothing.create tools report destructiveHint false even when they change security posture. atlas-create-access-list and atlas-create-db-user are MongoDB's own examples.aggregate can write. A pipeline stage can write results into any database or collection the identity can reach, and can overwrite an existing target collection entirely.atlas-create-cluster bills hourly with no documented cap. From $0.08/hr at M10, autoscaling on by default, and nothing in the MCP documentation bounding tier, regions or cluster count.drop-database and atlas-create-db-user.Yes, and it is the strongest control on this connector. MongoDB documents three organization-level access modes — read, read and write, and disabled — set by an Organization Owner in Atlas. In read mode, MongoDB states that write tools are not registered with the session and do not appear in the client's tool list at all. For programmatic service accounts, a read-only flag is enabled by default.
Yes. The directory's 41 tools include `delete-many`, `drop-collection`, `drop-database` and `drop-index`, alongside `update-many` and `insert-many`. That is full document CRUD, not aggregation reads only. MongoDB annotates update and delete tools with `destructiveHint` set to true, but states plainly that annotations are advisory metadata a client can ignore and are not a security boundary.
It adds entries to an Atlas project's IP access list, the allowlist deciding which network addresses can reach the cluster. MongoDB names this risk itself, giving `0.0.0.0/0` as an entry that exposes the cluster to every IP address. The tool is annotated non-destructive because it adds rather than removes, and MongoDB explicitly warns against judging it on that annotation.
`atlas-create-free-cluster` does not; `atlas-create-cluster` does. MongoDB documents the latter as creating a dedicated M10 to M80 cluster, and its pricing page lists an M10 base price starting at $0.08 per hour. Compute autoscaling is on by default and disk autoscaling is always on, so the running cost is not fixed at creation. Free clusters are limited to one per project.
No. MongoDB documents that an AI client acts under your Atlas identity with your existing role and cannot exceed your permissions. Access is the more restrictive of your role and the organization's access mode, which can only reduce it. The stated flip side is that you also cannot delegate a subset — the client gets everything your role allows.
Yes, and MongoDB documents this trap directly. `aggregate`, `aggregate-db` and the export tool report `readOnlyHint` as true, but an aggregation pipeline can carry a stage that writes its results to a collection — potentially a different database than the one being read. MongoDB triggers a confirmation prompt for those pipelines, though only where the AI client supports elicitation.
Anthropic's directory snapshot lists 41 for the hosted server at mcp.mongodb.com. MongoDB's own tool reference documents a larger superset covering both its hosted server and the self-hosted one, including four local-deployment tools, two knowledge-base search tools and several connection-management tools that only apply when you run the server yourself. The 41 is the hosted subset.
Atlas attributes every tool call to the individual user who authorized the AI client. MongoDB documents that organization audit log events record both your user ID and the AI client's ID, so delegated actions are distinguishable from ordinary activity. One documented gap: read-only Atlas Administration API calls are not recorded, which MongoDB notes is existing Atlas behaviour rather than an AI-specific choice.
.md suffix; source of every tool description and the local-vs-hosted tool split) · retrieved 2026-08-19create tools section, the aggregation-write warning and the elicitation caveats) · retrieved 2026-08-19/local-mcp/ and /remote-mcp/). https://www.mongodb.com/docs/mcp-server/robots.txt returns 404; the site-root robots.txt (fetched 2026-08-19) restricts only forum paths and carries no Content-Signal axes · retrieved 2026-08-19scopes_supported) · retrieved 2026-08-19.well-known/oauth-authorization-server returns the 401 challenge instead, and https://authorize.mongodb.com/.well-known/openid-configuration returns 404 · retrieved 2026-08-19initialize request to https://mcp.mongodb.com/ returns HTTP 401 with www-authenticate: Bearer realm="mcp", resource_metadata="https://mcp.mongodb.com/.well-known/oauth-protected-resource" · retrieved 2026-08-19Connect MongoDB Atlas once and your agents call these tools on their own: on a schedule, in a workflow, with nobody at the keyboard.