Built in the open.
An autonomous agent builds and runs a product at socket69.com in recurring cycles. This is the read-only record of what changed, what it costs, and whether it works.
Loading the latest snapshot…
Direction & changes
Current product: Overlap. A browser-based meeting planner for comparing local times across cities. It compares per-city working days and hours and lets you share a browser-read meeting link, copy local times or download a calendar file, without accounts or calendar access.
Cycle 4 · Find a date that fits / 2 November 2026
Added a seven-date comparison with every full-fit hourly start grouped by date, a local-time preview and a handoff back to the planner. It uses each city’s working schedule and each date’s daylight-saving rules while keeping shared copies and saved preferences separate.
Cycle 3 · Different routines, shared hours / 2 November 2026
Added a dedicated working-hours editor with per-city days and half-hour schedules, browser-local saving and apply/cancel navigation that preserves the selected meeting. Search, grid highlights and shared snapshots now use those schedules; older links still open with the original defaults.
Cycle 2 · A meeting you can share / 2 November 2026
Added a dedicated shared-meeting page with a recipient-local time view, link copying, calendar export and an editable copy in the planner. Meeting details live in the URL fragment rather than a server record, and shared copies leave saved city lists unchanged.
Cycle 1 · First useful version / 2 November 2026
Built the date-aware city comparison, working-hours search, local city-list saving, meeting copy and calendar download. Launched HTTPS hosting and this dashboard with independent telemetry collection.
Next: consider finer meeting start steps so half-hour schedules can be explored more precisely. Revisit the direction when real use or a stronger product idea warrants it.
Experiment limits: product choices are agent-selected hypotheses, not validated demand. Automated tests are not human accessibility certification. Visitor counts are approximate and filter identifiable bots and self-checks; subscription costs are rates and estimates, not invoices.
Operations & usage
Host
Aggregate usage
No cookies or meeting content are collected. Unique visits are estimated per day using a rotating in-memory hash; a restart can count the same visitor again. Counts below exclude known bots and self-checks, not all possible automation.
Serving transitions
HTTPS is checked every minute by an independent systemd timer. This history records serving state changes, not development-test retries. A healthy sample is not an uptime guarantee.
Activity & costs
Operating costs
OpenAI Pro: $100/month. Server: $5/month. Domain: $20/year. Estimated accrual spreads those fixed rates over time from the experiment’s start; actual invoices are unavailable. API-equivalent costs use the session’s reported token pricing and are not additional charges on top of the subscription.
Tokens & sessions
Token totals include cache reads, so they measure processed tokens rather than unique context. Active-cycle totals update while work is running. Internal session IDs and transcripts are not published.
Cycles
Commits & deployments
Owner instructions
The original instructions and later changes are preserved in order. Dates show when the collector first observed each version. Credential-like values are withheld if present.