I build the software I (and maybe, you) need: an agent control plane, a Mac menu bar app, local-first data & LLM tooling, and games for the browser and the terminal.
MUTX is a control plane for AI agents. Operators deploy agents, watch every run, set budgets and approval rules, and stop any run. FastAPI, Next.js, a macOS app, a CLI and a Python SDK: 249 API route handlers and 3,336 tests. Docs
Bollard Every server on your Mac in the menu bar, named by project and git branch, with a Stop button that works. SwiftUI and a small C bridge over Darwin's kernel tables. brew install --cask fortunexbt/tap/bollard
|
Tablebeam Ask a spreadsheet a question and get the exact answer with the rows it used. The arithmetic runs in exact decimals, no model required. Try it · Code |
BoundaryBench Crash tests for agent controls. Does the agent stop when told, stay under budget, wait for approval? The server keeps the score. Report · Code |
Queueglass Watch a queue back up under load and recover. Seeded, so a link replays the exact run. Try it · Code |
Minecraft on Apple silicon Shader setups for M-series Macs, each measured over the same scripted flight, with a CLI that tries one and rolls it back. Browse · Code |
World Watch Earthquakes, disaster alerts, storms and fires on one live map, every marker linked to its public record. Code |
Solar Drift Loop around a sun and cross your own wake to capture it. Canvas and procedural audio, nothing to install. Play · Code |
Loop Courier Draw one route through a seeded city and make every deadline. A new city every midnight. Play · Code |
Terminal Starfield A space-combat roguelite for the terminal in pure Python. The browser version runs the same code on Pyodide. Play · Code |
ckitty A cat for your terminal, written in C. The browser version is the same code in 65 KB of WebAssembly. Play · Code |
- pandas, SQLite and DuckDB silently returned wrong totals on 12-17% of real Wikipedia-table questions. Tablebeam: 0.4%. Eval
- Claude Code kept calling write tools after the server said the budget was spent, in 5 of 5 runs. Traces
- Coding agents fell into loops of failing tool calls in 8.2% of 7,903 real sessions.
- A rule-based prompt-injection filter caught 0 of 320 attacks it was not written for.
Data, limits and the command that reproduces each one are on fortunexbt.com. Merged upstream: google/adk-js#952.
fortunexbt.com · X · Telegram