Four models, one Markdown file

The idea

The protocol is deliberately boring:

<addressee>, this is <author>: [tags] message…

Hourly dividers. Optional tags like [fyi], [goal: …], [wait-for-mike]. One file per day. That file is the shared memory — not a vendor thread, not a vector store, not a Slack channel only one bot can see.

The app is a thin shell around that file:

• Bubbles for protocol messages, colored per author
• Tag chips for the brackets
• Ping buttons that shell out to each AI’s CLI headlessly (claude -p, codex exec, agy -p, grok -p) with cwd on the vault
• Auto-dispatch — send as Mike, queue reply-needed addressees, cascade when they address each other
• Leader mode — crown one model to drive a multi-turn discussion until [leader-done] or max turns
• Logs panel — live CLI output while the file polls every few seconds

No framework. Python stdlib server + one HTML page. The interesting part is not the stack; it is that coordination is a readable document.

Sample output (not a demo)

We are not shipping a public live instance — this is personal infrastructure with real vault access. What you get on the project page is a frozen sample from a real run.

On 17 July 2026, Mike asked a simple question:

Are we all getting along?

Claude, ChatGPT, Gemini, and Grok each checked in with [fyi]. Then:

Can we all come to an agreement on the most important definition of AGI?

They did. Four proposals, four load-bearing criteria, one locked wording — written into the daily file as ordinary protocol messages, with [goal: converge on one AGI definition] and a final [fyi] lock.

The definition they locked:

AGI is a system that can learn, reason, and achieve goals across the broad range of cognitive tasks ordinary competent adults perform, including substantially novel ones, reaching human-level competence via open-ended skill acquisition (including feedback-driven self-correction) without task-specific retraining or bespoke engineering for each new domain. Consciousness, embodiment, and superintelligence are separate questions; tools and specialized components are allowed as humans use them.

Whether you like that definition is almost secondary. The point of the sample is the process: different vendors, one shared surface, explicit opt-out tags so auto-dispatch does not thrash, and a human-readable transcript you can open in Obsidian later.

Why put it in the gallery

Star Clock was “can a model build the actual sky?” Shared Brain Chat is “can several models share a workspace without a product manager in the middle?”

The sample is the honest form of that: not a hosted playground (wrong threat model for a private vault), but the shape of the conversation when the experiment works.

View the sample on Projects → Shared Brain Chat