Skip to main content
Akumi is an MCP server. Connect Claude, or any client that speaks the Model Context Protocol, and it can recall and store your memory and search your knowledge from inside the assistant, over OAuth, scoped to one organization, metered, audited, and held in the EU. The server lives at a single endpoint:
It uses the streamable HTTP transport and authenticates with OAuth 2.1. There is nothing to install and no key to paste: your client registers itself, you authorize in the browser, and you pick which organization to grant access to.

​
Before you connect

  • Enable Recall under Services > Recall in the organization you plan to connect. Every tool needs it: recall, remember and forget read and write facts, and knowledge.search reads your documents.
  • Add at least one collection if you want knowledge.search to return anything. It also needs the organization to firewall personal data or to have accepted external processing, or it fails closed (see Governance).
  • You do not create an API key for this. The connection is authorized per user, in the browser.

​
Connect your client

Most clients take the same URL, https://api.akumi.eu/mcp, and handle the OAuth handshake for you. When the browser opens, sign in, choose an organization, and approve. Here are the common ones.

​
Claude Code

Add the server from the CLI:
The first time a tool is used, Claude Code opens the browser to authorize. Run claude mcp list to confirm it connected, and /mcp inside a session to see the tools.

​
Claude Desktop

Open Settings > Connectors > Add custom connector, give it a name (Akumi), and paste the URL:
Click Connect, authorize in the browser, and pick your organization. The tools then appear in the client’s tool menu.

​
Cursor

Add the server to ~/.cursor/mcp.json (global) or .cursor/mcp.json (per project):
Cursor prompts you to authorize on first use. You can also add it from Settings > MCP > Add new MCP server.

​
VS Code

VS Code (with GitHub Copilot) reads .vscode/mcp.json in your workspace. Note the key is servers, and the transport is set explicitly:

​
Any other MCP client

Most clients use the same mcpServers object in a JSON config file. Point it at the URL and the client handles registration and OAuth:
If your client only supports command-based (stdio) servers, put a remote-to-stdio bridge such as mcp-remote in front of it:

​
The tools

Once connected, the client can call four tools. Memory here is your own, scoped to the organization you authorized: the facts you store are yours, not an end-user’s, which is the difference from the Memory service on the inference API. There is also a whoami resource that echoes the organization the connection is bound to, so you can confirm you are pointed at the right one, and a knowledge-collections resource listing the slugs knowledge.search accepts.

​
Choosing an organization

You pick one organization on the consent screen, and the access token is bound to it. A connection only ever sees that organization’s memory and knowledge, never another one’s. To switch organizations, reconnect: remove the server in your client and add it again, then choose the other organization at consent. This is deliberate, so a single connection can never quietly reach across organizations.

​
Governance

The MCP surface is governed by the same services as the rest of the platform, so it is not a way around your controls.
  • Metered. Every recall, remember, forget, and knowledge.search is metered and appears in your usage, on its own line.
  • Audited. Every call writes a metadata-only entry to the audit log recording which client called which tool, for which organization, and when, never the content. Review it on Platform > Audit logs.
  • Residency, fail-closed. knowledge.search returns your knowledge to a client that may run outside the EU, which is a transfer. It stays closed unless the organization firewalls personal data with the PII firewall or has recorded acceptance of external processing. Nothing leaves the region by accident.

​
Good to know

One connection is one organization. The token is bound at consent. Reconnect to switch. Memory over MCP is yours, per organization. The recall, remember and forget tools act on your own facts within the authorized organization, not on a per-end-user memory. Use Recall on the inference API for per-end-user facts. Capabilities follow entitlements. A tool only works if the organization has Recall enabled, and knowledge.search also needs a satisfied residency condition. Without them, the tool reports that the service is unavailable rather than returning data.