Demos · shipped in public
Short demos and prototypes from the last few months. More on X.
Evals built on a simple belief: everyone compares models, almost no one compares harnesses. Same model, same task across codex, pi, opencode, cursor, devin, and claude — measuring solve rate, $/solve, and latency.
Save repeatable workflows as computer-use skills. After the agent finishes a workflow, tell it to save_skill and just call it next time — or teach it one by recording. Works with Claude and any agent that speaks MCP.
A computer use MCP that works with any agent: Claude, Devin, Grok, whatever you run. Lets the agent work across your computer in the background, verify itself in a feedback loop, and take on the annoying tasks you'd rather not.
A yolo /remote-control built with /goal. It spins up a tiny server on the laptop, mints a fresh token and QR code, and lets you drive a Codex session from your phone.
Codex’s new /goal command running unattended. Set a one-line objective, walk away, come back to working games. The thread has notes on /goal resume and computer use self tests.
Found an open-source computer use extension for Pi on X, had my agent audit it for security, then patched pi-gui to support it and added a built-in terminal. v0.1.0-beta.26.
From “beta is around the corner” to v0.1.0-beta and counting. Incremental polish on the way to a real release.
A Pi extension that lets the agent generate 3D worlds, wiring a creative tool straight into the agent loop.
Another step toward the Codex-app feel for Pi. Collapsible tool calls, an in-timeline diff viewer, and tighter session UX.
Built a Codex/Claude style subagents flow for Pi in a few hours. Spawn parallel agents, message and wait on them, non-blocking orchestration.
Building the Codex-app experience for Pi. Native macOS chrome, multiple parallel sessions, a streaming timeline with inline diffs, slash commands, git worktree lifecycle.