Introducing the MCP server

Akumi is now an MCP server. Connect Claude Desktop, Claude Code, or any client that speaks the Model Context Protocol, and it can recall and store your organization's memory and search your knowledge, over OAuth, scoped to one organization, metered and audited, and held in the EU.

Your team already works inside AI clients. They are most useful when they know your context: the decision you made last quarter, the house fact that never made it into a doc, the page that answers the question someone is about to ask again. The problem is how that context gets in. The easy path is to paste it into the chat, or to wire up a memory server that hands your data to whatever connects, wherever it happens to run. For a team with regulated data, that is an unlogged transfer waiting to happen. You want the assistant to remember. You cannot afford for it to remember somewhere you cannot see.

What we built

The Model Context Protocol is the open standard that Claude and other clients use to reach external tools. We implemented it as a server inside the same platform that already runs your inference, so connecting an AI client to your governed memory is a standard OAuth flow, not a custom integration.

Once connected, a client sees four tools:

• recall reads a fact from your organization's long-term memory.
• remember stores a new fact.
• forget removes one.
• knowledge.search answers from your EU-resident RAG collections.

These are not a new copy of your data. They are the same governed Memory and RAG services the API already uses, reached over a different protocol. Metering, audit, entitlements, and residency all come along for free, because they live in those services, not in the surface on top.

Governed the same way as everything else

Connecting is Dynamic Client Registration plus OAuth 2.1. The user authorizes on a consent screen that names the transfer in plain language, picks one organization, and the token is bound to it. A client only ever sees that organization's memory and knowledge, never another tenant's. Switching organizations means reconnecting, on purpose, rather than a single token quietly reaching everything you can touch.

Every call is metered on its own line and written to the audit trail, with the client and the acting user recorded. When someone asks which tool ran, for which organization, and when, you answer with the log rather than a guess.

Knowledge egress fails closed. knowledge.search stays shut unless the organization firewalls personal data with the PII firewall or has accepted external processing on the record. Returning your knowledge base to a client that runs outside the EU is a transfer, and we treat it like one, so nothing leaves the region by accident.

Connect in one command

If you use Claude Code, adding the server is a single line:

claude mcp add --transport http akumi https://api.akumi.cloud/mcp

Authorize in the browser, choose your organization, and the tools appear in the client. Claude Desktop and other MCP clients take the same URL through their own connection settings.

Where this goes

The MCP server is the first piece of a larger direction: Akumi as the governed, EU-sovereign knowledge layer your AI clients connect to, with more connectors to follow. It reuses the governance we already built for the API, so the safe path and the convenient path stay the same path.

Create a key, enable Memory, and connect your client in a couple of minutes.