Cloud MCP

Connect Claude, ChatGPT, Claude Code and other MCP clients to your Cognee Cloud workspace

Cognee Cloud speaks MCP (Model Context Protocol) in two ways:

• The hosted MCP server at https://mcp.aws.cognee.ai/mcp. Nothing to install: you add the URL to your client and sign in with your Cognee account. Any client that supports remote MCP servers with OAuth can use it, including Claude (web and Desktop), Claude Code and ChatGPT.
• Running cognee-mcp yourself with an API key. For clients that can only launch a local server, or when you want to pin a version and control the environment. The Integrations page in the Cloud UI generates these configs.

Both give your client the same tools, and both act on your Cognee Cloud workspace. This page covers the hosted server first, then the self-run option.

Connect to the hosted server

Claude Code

Register the server once for your user, then sign in from inside Claude Code:

claude mcp add --scope user --transport http cognee https://mcp.aws.cognee.ai/mcp

If Claude Code is already running, restart it so it picks up the new server. Then run /mcp, select cognee and choose Authenticate. Your browser opens Cognee’s consent page: click Allow Access once, then sign in with your Cognee account. Back in Claude Code, /mcp shows the server as connected.

Claude (web and Desktop)

Custom connectors are available on every Claude plan (a Free plan can have one), in Claude on the web, Claude Desktop and Cowork.

1. Open Customize → Connectors and choose + Add, then Add custom connector.
2. Name it Cognee and enter the remote MCP server URL https://mcp.aws.cognee.ai/mcp. Leave the authentication options at their defaults (Sign in now, no request headers).
3. Choose Add. On Cognee’s consent page click Allow Access and sign in with your Cognee account.

On a Team or Enterprise plan, an Owner adds the connector under Organization settings → Connectors; each member then connects it from Customize → Connectors with their own Cognee account. Anthropic’s custom connectors article has the per-plan details.

In a conversation, enable the connector from the tools menu if it is not on already.

ChatGPT

ChatGPT connects to remote MCP servers through its developer mode. Connectors you create are tied to your ChatGPT account, so the web and desktop apps share them. Menu names differ a little between plans and change often; this is the current shape.

1. Turn on developer mode. On individual plans it is under Settings → Apps & Connectors → Advanced settings → Developer mode; on Business and Enterprise workspaces an admin enables it under Workspace settings → Apps.
2. On the Apps & Connectors page choose Create. Enter the name Cognee, the MCP server URL https://mcp.aws.cognee.ai/mcp, and choose OAuth as the authentication. Confirm that you trust the application.
3. Choose Create. ChatGPT scans the server’s tools and asks you to sign in: click Allow Access on Cognee’s consent page and sign in with your Cognee account.

In a chat, open the tools menu and enable the Cognee connector for that conversation.

Other clients

The hosted server works with any client that supports MCP over Streamable HTTP with OAuth 2.1: the client registers itself (dynamic client registration or client ID metadata documents) and uses PKCE. Give the client only the URL:

https://mcp.aws.cognee.ai/mcp

The client discovers the sign-in endpoints on its own. The hosted server does not accept API keys. For a client that only supports API keys, or that can only launch a local server, run the MCP server yourself.

Try it

Ask your client to store something, read it back, and clean up:

Use Cognee to remember in the dataset mcp_test: “The hosted MCP connection works.”

Use Cognee to recall from mcp_test: what works?

Use Cognee to forget the dataset mcp_test.

The first answer reports Stored permanently in knowledge graph (dataset=mcp_test, status=…). Storing runs the full ingestion pipeline, so it can take from a few seconds to a few minutes depending on the amount of text. The dataset also appears on the Datasets page of the Cloud UI.

What happens when you connect

Sign-in. The hosted server is an OAuth authorization server towards your client and signs you in through the same Cognee account you use for the Cloud UI. The client receives a token that works only with the hosted MCP server; it never sees your Cognee credentials or session. A connection stays signed in while you keep using it. After 30 days without use, or a year after you connected, the client asks you to sign in again.

Which workspace. Every call goes to your current workspace, the one the Cloud UI opens when you sign in. With no workspace selected, the server uses your only workspace, or else the only one you own; if that still leaves several, the first call tells you so and asks you to switch to one in the Cloud UI. The server remembers your workspace for 5 minutes, so after you switch workspaces in the Cloud UI, a connected client can keep using the previous one for up to 5 minutes. Nothing lets a client reach a workspace you are not a member of: the server passes your own sign-in token to your workspace, which checks it like any other request.

Datasets. Unless you name a dataset, each client stores into a dataset of its own, named after the client: claude_code_memory for Claude Code, anthropic_claudeai_memory for Claude. Default Dataset Resolution explains the rule. recall without a dataset searches everything you can access, so memories written by one client are visible to the others. To use a specific dataset, say so in the conversation (“remember this in the dataset project_x”) or in the client’s instruction file.

Sleeping workspaces. A workspace that has been idle for a while goes to sleep. The first MCP call wakes it: the hosted server waits for the workspace to start, so the call usually succeeds, but it can take about a minute. Later calls are fast.

Credits. Calls made through MCP are billed to your workspace exactly as the same operations from the Cloud UI or the API.

Tools

The hosted server lists all four Cognee tools directly. There is no search_tools / call_tool pair to go through, unlike the default mode of a self-run server.

Tool	What it does	Writes?
remember	Store text or a file (up to 10 MB) in memory: add and cognify in one step.	Yes
recall	Search memory. Without a dataset, searches everything you can access.	No
forget	Delete one item, one dataset, or everything you own.	Yes, destructive
cognify_status	Progress of ingestion started with remember(background=true).	No

Parameters and defaults are in the Tools Reference.

Troubleshooting

What you see	Why	What to do
Authorization session mismatch. Please try authenticating again.	The consent form was submitted twice; the first submit already went through.	Start the connection again and click Allow Access once.
Invalid or expired transaction	The sign-in took longer than 15 minutes, or belongs to an attempt that already finished.	Start the connection again.
You don't have a Cognee workspace yet.	Your account has no workspace.	Sign in at platform.cognee.ai and create one, then retry.
You belong to several Cognee workspaces (…) and none is selected.	No workspace is selected, and you belong to several without owning exactly one of them.	Switch to the workspace you want in the Cloud UI, then retry.
The first call after a quiet period takes a minute, or times out.	Your workspace was asleep and is starting.	Wait a moment and retry; the workspace stays awake while you use it.
Tools fail with an authorization error after working before.	The client’s token was revoked or has expired.	Disconnect the connector in the client and connect again.

Run the MCP server yourself

Use this when your client only launches local (stdio) servers, only supports API keys, or when you want to pin the cognee-mcp version. You need the API Base URL and an API key from the API Keys page.

Client launches the server (stdio)

This is the form the in-product Integrations page generates. It runs cognee-mcp through uvx (requires uv); the first launch downloads the package and its dependencies. The version is pinned because a bare cognee-mcp runs whatever version uvx cached on that machine; the Integrations page pins the version Cognee Cloud has been tested with.

{
  "mcpServers": {
    "cognee": {
      "command": "uvx",
      "args": ["cognee-mcp@0.5.6"],
      "env": {
        "COGNEE_BASE_URL": "https://your-tenant.aws.cognee.ai",
        "COGNEE_API_KEY": "your-api-key"
      }
    }
  }
}

Copy the API Base URL exactly as shown on the API Keys page: the tenant-… label in it is how cognee-mcp recognizes a Cloud workspace and authenticates with your API key. For per-client config file locations, see Integrations.

Server runs separately (HTTP)

Start the server pointed at your workspace, then give the client its local URL:

cognee-mcp --transport http --port 8001 \
  --api-url https://your-tenant.aws.cognee.ai \
  --api-token your-api-key

{
  "mcpServers": {
    "cognee": {
      "url": "http://localhost:8001/mcp"
    }
  }
}

--api-url / --api-token put the server in API mode, so every tool call is sent to your workspace. Cognee Cloud & MCP explains API mode, authentication schemes and the difference from --serve-url.

A self-run server offers the same memory tools; in its default mode it lists remember, recall, forget and code_search and reaches the rest through search_tools / call_tool, as described in Tool Modes.

Next steps

Tools Reference

Parameters, defaults and dataset rules for every Cognee MCP tool.

Workspace Lifecycle

How a workspace sleeps and wakes, and what that means for the first call.

Cloud SDK

Connect to Cognee Cloud programmatically using the Python SDK.

Integrations in the Cloud UI

Generated configs for clients that run the MCP server locally.