Builds Agents That Ship

A terminal coding agent written from scratch in TypeScript: its own agent loop, its own ANSI renderer — no Ink, no React, no curses — sessions, subagents, skills, hooks and MCP. It reads a codebase, reads a screen, and drives a browser through a semantic snapshot of the live DOM rather than a pixel guess; on Windows it operates the desktop as well. Shell commands run under OS confinement: bubblewrap on Linux, Seatbelt on macOS, deliberate policy and consent on Windows. An unattended run has to prove what it did — an application counts only when a tool watched the site's own confirmation appear, for that vacancy.

• 51,870 lines
• 237 modules
• 4,889 tests
• Linux · macOS · Windows
• 7 runtime dependencies

An autonomous personal agent running unattended on a single host since December 2025. An original ReAct runtime, 65 tools, 12 swappable model roles, hot structured memory over a vector archive, and approval gates on anything irreversible. It triages mail, briefs me each morning, supervises a live trading bot — arithmetic is always re-derived in Python, never trusted to the model — starts and steers remote coding-agent sessions over tmux, and edits its own source.

• 18,353 lines
• 65 tools
• 12 model roles
• 16 scheduled jobs

A nutrition assistant people pay for. Photograph the plate, say it out loud, or type one line — the model reads it and records calories and macros against a history it remembers, so the answer on day ninety knows about day three. Nine fully translated languages, an installable React 19 PWA, FastAPI behind Caddy on a Hetzner host, MongoDB Atlas and Firebase identity, Stripe subscriptions.

• 29,741 lines
• 2,887 checks green
• 97% server coverage
• 9 languages

A community that remembers — a place, not a feed. An invite-only platform with a memory-first architecture: a living Space instead of a timeline, and a permission-aware AI layer that answers only from sources the asker is entitled to see, and refuses rather than inventing one. Access is resolved before retrieval scores a single term, because searching first and filtering the output afterwards has already leaked the thing it was meant to hide.

• 31,681 lines
• 110/110 permission cells
• 20/20 refusal cases
• 55 verification scripts

A growth-experimentation engine: a reinforcement loop for marketing. It forms hypotheses, runs them on real channels, measures clicks and conversions, and learns through a Thompson-sampling bandit. The model never decides a number — every reward is computed in code from the database, a retried delivery cannot count one conversion twice, and how much autonomy a channel gets is product policy rather than a line in a prompt.

• 16,093 lines
• 375 tests
• 7 channels
• Full, approval, suggest-only

A Chrome extension and meeting-analysis service for Google Meet: live English-to-Russian caption translation, audio-based recovery when captions arrive incomplete, canonical transcripts and search. Meet offers no caption API, so the region is recognised by what it is and never by where it sits — Chrome reports zeroed rectangles for background windows, and a rule written around position went blind the moment somebody looked away.

• Manifest V3
• Tab audio only
• Never the microphone

A rigorously honest monitor for hidden correlations across twelve independent real-world systems — crypto, earthquakes, space weather, hardware quantum randomness, the solid-earth tide. Adaptive anomaly detection, shuffle tests, out-of-sample backtests, block bootstrap. No AI anywhere in the analysis. The finding so far is that no cross-domain edge survives out of sample, and that gets published too.

• 12 live streams
• 140 tracked patterns
• MIT

A manufacturing test-yield dashboard built to a written specification: FastAPI and SQLite, a three-panel dashboard, and a plain-English chat with tool calling where every figure is computed in Python against the database, never by the model's own arithmetic. A Selenium test drives the real browser and asserts the 60% yield scenario end to end.

• Selenium end to end
• Grounded tool calling
• MIT