Connections
OAuth-based, consent-first. Users authorize once; credentials live in a per-device wallet — not a config file, not a running process.
What we believe
The model is the engine.
Nobody buys an engine.
“Apple didn’t sell music players. They sold a thousand songs in your pocket. Stripe didn’t sell payment APIs. They sold seven lines of code instead of seven weeks of integration. The companies that win in any new technology layer are the ones that hide the technology and surface the experience.”
Right now, to get an AI assistant to read your calendar, you install a server, edit a JSON file, and hope a background process stays alive. To connect a second tool, you do it again. To share the setup with a colleague, you write documentation. We have built models that can reason across thousands of pages of legal text — and we ask their users to be sysadmins.
This is the wrong layer. AI’s bottleneck is no longer model intelligence. It is integration. The wire is solved — MCP is now an AAIF-governed industry standard, co-owned by Anthropic, Block, OpenAI, and Google. The experience around the wire is not. That is where Ferridis lives.
The layer that hides the wires wins. Ferridis is built to be that layer — open, well-named, properly attributed.
Context
The standard
is settled.
The layer above isn’t.
In December 2025, Anthropic donated MCP to the Agentic AI Foundation — a Linux Foundation directed fund co-founded with Block and OpenAI. Google’s A2A protocol joined the same foundation. The wire protocol for connecting AI to tools is now an industry-coalition standard, not anyone’s private project. That is good news.
It also means the interesting work has moved up the stack. Connections instead of servers. Federated discovery. Lazy schemas. An intent vocabulary. A coherent fallback to the browser. That is where Ferridis lives — not as a replacement for the standard, but as the opinionated layer above it: a reference design and a Rust implementation for what the experience around the wire could look like.
Architecture
Six layers.
One user experience.
The user sees one thing: I connected my calendar. The assistant uses it. Underneath that simplicity, six layers do the work.
Capability mesh
Federated discovery with two-stage advertisement: a compact manifest always in context, the full schema loaded only when the AI actually needs it.
Web-native specs
OpenAPI 3 and AsyncAPI — no new spec language to learn. Every modern service already publishes one of these, or can with minimal effort.
Bidirectional channels
The world pushes back. SSE and WebSocket channels let services send events to the AI — not just respond to it. Push is first-class, not bolted on.
Intent layer
Goals, not tool calls. High-level instructions route to the right capability across the mesh — the model doesn’t memorise 500 tool descriptions.
Browser fallback
Anything not yet wired up natively falls back to a browser session — a universal escape hatch that lives inside the same protocol.
About the name
Alexandros Ferridis d. 8 November 2012
Ferridis is named for Alexandros Ferridis, the high school teacher in Greece who shaped how the project’s author thinks about work and about the people he does it with. He passed away from cancer on the 8th of November 2012.
He was the kind of teacher whose influence you only fully understand much later. He didn’t just teach a subject — he taught a way of approaching problems carefully, and of treating other people’s intelligence with respect. The instinct to share what you build rather than hoard it, to work in the open rather than behind closed doors — that came from him.
This protocol is built on a belief that good infrastructure should be generous: open, well-attributed, useful to anyone who needs it. That belief is his. Naming it Ferridis is the only way to keep his name moving through a world he isn’t in anymore.
— George AndrikopoulosThe work begins now.
If you build for AI, integrate AI, or use AI seriously enough to feel the friction — follow along. Push back. Help shape it. The design process is public.
View on GitHub — George Andrikopoulos