A product people paid for

Captain was an always-on chief of staff that took real actions instead of drafting them. It read incoming email, drafted replies in the user's voice, scheduled meetings against a real calendar, and sent nothing without approval on anything that could embarrass or commit the user. It found its buyers among account executives, who used it for pre-call briefs. 140 people used the beta. 90 percent of the actions it proposed were accepted as-is.

What it did

A multi-agent orchestration in TypeScript on Mastra, with specialized sub-agents behind tool integrations for Slack, Gmail, and Calendar, and a voice front end for people who talk to their phone between meetings.

The part that earned money was the brief. Before each call, a retrieval pipeline over the user's email threads, Slack history, prior meetings, and a memory store pulled everything relevant into one page: who is on the call, what was promised last time, what is open, what to ask. The pipeline was pgvector and OpenAI embeddings. It was not clever. It was fast, and it was right often enough that people stopped preparing calls by hand.

Production-LLM discipline, early

This was the first product where I ran the disciplines that every later system inherits.

• Every proposed action instrumented and traced, so a wrong send could be explained after the fact.
• Approval gates on high-risk actions. Send versus draft was a policy decision, not the model's call.
• Structured outputs for every tool call, so reliability came from the schema rather than from the prompt.
• Prompt and guardrail changes evaluated against real accepted-versus-rejected outcomes, not against vibes.

The 90 percent acceptance number is the eval. It is the share of proposed actions users approved without editing. It is also the bar I still use for "good enough to ship."

What it taught me

Captain proved that people will pay for an agent that acts, and that they will only keep paying if it never acts wrongly in public. The approval gate was not a compromise on autonomy. It was the product. Everything I have built since, Greenlight, Carrier Desk, Meridian OS, FieldMate, is a more rigorous version of that one idea: the model proposes, and a deterministic layer decides.

The honest gaps

• The retrieval pipeline is described here, not shown. Captain's data was users' private email, so the public repo carries the architecture and none of the corpus.
• 140 users is a beta, not a business. I stopped in April 2026 to found Toglo.
• The public repository is the early version. The monetized version lived in a private repo with real user data.