Spotify Deleted Its AI Platform and Nobody Noticed

Spotify’s MCP Gateway

Spotify describes how it built an internal Model Context Protocol (MCP) gateway to support employee-created AI tools, then replaced much of its custom implementation with open-source components without disrupting users.

Scale and adoption

The gateway handles about 163,000 tool calls per day, serves roughly 150 MCP servers, and has grown from around 50,000 monthly calls to about 4 million per month. The speakers say all Spotify employees use it.

Why Spotify built it

During an early wave of MCP experimentation, employees created servers using inconsistent technologies and ad hoc authentication. Spotify’s existing backend frameworks did not yet support MCP, so the company built a gateway to make deployment easier and safer.

Building on existing systems

Spotify has thousands of gRPC services. Its first gateway let teams expose those services through MCP by transcoding MCP’s JSON-RPC traffic into gRPC. This enabled teams to create MCP servers using familiar tools and led to rapid adoption.

Moving to open source

The custom gateway accumulated features and code owned by different teams, making maintenance, deployment, and ownership harder. As Envoy added a native MCP filter, Spotify moved toward a more standardized, Envoy-based architecture.

The gRPC transcoding functionality moved out of the gateway and into a library. The underlying capabilities remain, but the platform became simpler and more open.

Authentication and security

The gateway is the controlled entry point to MCP servers. A small YAML configuration specifies where traffic goes and what token permissions are needed.

For human-driven agents, Spotify uses OAuth token exchange (RFC 8693) to create short-lived, narrowly scoped tokens just before a request. Lower-level credentials are kept out of the language model’s context, and the MCP server receives the user’s identity and only the permissions it needs.

Account linking and autonomous agents

Users link accounts and grant permissions for a server, after which access can generally proceed automatically.

For headless or scheduled agents, Spotify assigns the agent its own identity and permissions, with the same gateway policy checks and audit logging—rather than having it impersonate a person.

Enterprise authentication

The speakers discuss an emerging enterprise authentication management approach intended to reduce repeated login and consent prompts. They note that gateway support remains a challenge because the mechanism currently works for only one node, and that account linking is still required in their setup.

Key lessons

• Build on existing systems and standards.
• Keep interfaces narrow so implementations can change behind them.
• Make secure usage the easiest path.
• Replace custom components with mature open-source solutions when they become available.

Reviews, guides, or tutorials

No product review or step-by-step tutorial is presented. The talk is an experience report on Spotify’s MCP architecture, security model, and migration.

Main speakers and source

Janek and a second Spotify employee, whose name is transcribed inconsistently as Rinut/Reynold, are both identified as Spotify employees.