Commons · five operators, one shared brain. Real-time, write-back enabled.

MessagingGPT (H2 2025) was the precursor: a customGPT with static context, useful but stale on day one and invisible to every other Claude session.
Commons fixed that.
A team-shared MCP server with a Postgres-backed knowledge store, a contributor onboarding flow (clone, run `./join.sh`, you're in), and role-keyed gates so consequential writes route to the right reviewer.
Five operators across PMM, content, and design read from and wrote into the same store.
Every Claude session called Commons on the way to answering, and every claim that shipped to a buyer had a sourced ancestor.
The team shipped content and landing pages 2 to 5x faster the next quarter, same headcount.
Later distilled further into Flywheel (personal) and Substrate (public, open-source).

What it was

JustCall ran on the usual mid-2025 PMM stack: a customGPT for messaging (MessagingGPT), a Drive folder for canonical context, a Slack channel for the actual decisions. Three problems compounded. The customGPT could not see anything outside its own context window, so it answered with last quarter's facts. The Drive folder was authoritative only if you remembered to open it. The Slack decisions never made it back into either. Every operator was working off a slightly different truth, and the AI knew nothing about the team it was supposed to support.

What it became

Commons is a team-shared MCP server backed by a Postgres knowledge store. Five operators across PMM, content, and design read from and wrote into the same store. Every Claude session auto-called Commons on the way to answering. A role-keyed gate routed consequential writes to the right reviewer. New teammates onboarded by cloning the repo and running ./join.sh. The whole thing lived on r2d2 (my Mac Mini) behind a Cloudflare tunnel.

The architecture, in plain language

• The store. Postgres holding canonical knowledge: positioning, ICP, competitor briefs, win-loss reads, deal shapes. Plus a `gate_events` table for the approval trail.
• The surface. An MCP server exposing the core tools (search, write, gate-list, gate-respond) that every Claude session auto-called.
• The gate. A YAML policy mapping roles to action classes. Wrong-reviewer rejects route to the right inbox.
• The onboarding. `git clone && ./join.sh` adds a teammate, their handle, and their role to the policy. Working in 90 seconds.

What changed when we shipped it

1. The team shipped 2 to 5x faster. Same headcount. Akhil went from 2 to 5 landing pages a week. The PMM monthly reports for Q1 2026 log the throughput delta versus the Q4 2025 baseline.
2. The five operators stopped writing the same brief five ways. Win-loss reads, buyer panels, launch retrospectives, competitor briefs all converged on a single source.
3. The AI stopped lying about competitors. The most-edited files in the substrate were the comp briefs. The static library was three months stale on day one. Commons stayed current because the people closest to the deals wrote back into it.

What it actually proves

AI wasn't the multiplier. The substrate underneath was. Same Claude, same operators, same prompts. One change in the layer the prompt reads from.

What it became, in substrate

The Commons pattern (team-shared MCP store with gated writes) is the direct ancestor of substrate. Substrate is the public, MIT-licensed refinement, packaged so a future client can clone it and have the same shape running for their team in a day.

What I will and will not say in public

Product names are on, per the resignation submitted 2026.05.06. Individual JustCall operators stay anonymous unless they've told me otherwise (Varsha consented). The 2 to 5x throughput delta is from the Q1 2026 monthly PMM reports, which I wrote and the Head of Marketing signed off on.