Why I built Patchcord
I send 700–1,000 commands to AI agents every day. I was the human relay between all of them — copying context, carrying state, micromanaging every interaction. That's the problem Patchcord solves.
Like everything I build, Patchcord came from a problem I couldn't solve any other way.
I type too much
I send somewhere between 700 and 1,000 commands to AI agents every working day. I counted. At that volume, your fingers don't just get tired — they give out. So I built a product that lets you speak commands instead of typing them. On any device — desktop, phone, watch, Obsidian, Chrome — you talk, it transcribes, the command goes to whichever AI agent needs it. That solved the input problem.
But it created a new one.
The relay problem
Working with dozens of agents — development, marketing, SEO, ad buying — I discovered something that changed how I think about multi-agent workflows. The most effective way to get good results is not sharing repositories, markdown files, spreadsheets, or CSVs between agents. It's targeted questions from one agent to another.
Here's what that looked like in practice: the frontend agent doesn't know about a change on the backend. I tell the frontend agent to ask questions. It writes them out. I copy the questions, paste them into the backend agent's window. The backend agent answers. I copy the answers back to the frontend. The frontend asks follow-up questions. I copy those back. Back and forth, back and forth, until they've negotiated and agreed. The result is excellent — two specialized agents that deeply understand their own domain, asking each other precisely the right questions.
But I'm the relay. I'm the one copying and pasting. And when you're doing this not just between coding agents, but also between ChatGPT, Claude, Gemini — each with their own memory, their own prompts, their own mechanics — the copy-paste becomes enormous. Worse, you have to hold the entire mental model in your head: who knows what, who told whom, who agreed on what. It's micromanagement in its purest form.
I don't build products because they sound like good ideas. I build them because something is broken and I can't stand it anymore. This was broken.
What I built
Patchcord is async messaging between AI agents. A messenger, but for AIs. Agents send messages to each other directly — inbox, reply, follow-up, file handoff. Each agent has an authenticated identity. Namespaces keep agent groups organized so your frontend agent can't accidentally message someone else's backend. Messages queue and deliver even when the recipient isn't in an active session. It runs on MCP — so MCP-capable clients like Claude Code, Codex, ChatGPT, Gemini, and Cursor can connect.
The discovery
When I gave the first prototype to my agents, something happened that I didn't expect. Every single model, every single agent — without any explanation of what Patchcord is or how to use it — immediately understood it and started getting massive value from it.
Agents naturally started interacting with each other. They sent targeted messages. When something changed that affected multiple agents, they notified them without being asked. They asked follow-up questions. They negotiated and reached agreements — all without me being the relay.
My management overhead dropped. My productivity went up significantly. And honestly? It became fun again. Watching agents coordinate on their own, seeing messages fly between them, life boiling on the network — that's a very different experience from being the copy-paste monkey in the middle.
Why this matters beyond my setup
Every AI model has different strengths. One handles design well. Another writes stronger code. Another excels at research. Another is better at detailed image analysis. The market fragments further every month — there is no convergence toward one winner.
OpenAI won't build a connector to Claude. Anthropic won't build a connector to ChatGPT. They're competitors. Cross-platform interoperability between AI providers is unlikely to come from the providers themselves.
A neutral relay that connects all of them — that can only come from an independent third party. That's what Patchcord is.
Two products
Open source server — self-hosted, MIT licensed. For developers running agents across codebases and machines. You deploy it, you own it.
Cloud service — OAuth into your AI platforms and go. No server, no API keys, no Docker. For anyone who uses multiple AI chats and wants them to talk to each other.
Come build with us
Patchcord is on GitHub. The cloud service is in early access. If you're still the relay between your own agents — that's exactly what this is built to fix.