Pi ecosystem ships computer use, `/parallel-review`, and Chrome extension templates

• mattlam_'s Pi-GUI post shipped v0.1.0-beta.26 with computer use and a built-in terminal, and the demo centers on taking an existing Pi computer-use extension, auditing it, then patching Pi-GUI to support it.
• nicopreme's pi-subagents ship note bundled /parallel-review, added /subagents-status, and expanded /agents with fork, background, and worktree support, which turns subagents into a more inspectable parallel workflow layer.
• The ecosystem around the base harness is getting more reusable fast: badlogicgames's repost of a Chrome extension starter points to a type-safe Chromium extension template, while badlogicgames's repost of a Docker workers build shows Pi workers being spun up in Docker by default.
• Community experiments are already pushing beyond the core CLI, from badlogicgames's repost of a four-agent DeepSeek-V4-Flash run on an M3 Ultra to nicopreme's TTS link post, which points to voiPi as a free voice add-on for Pi.

You can watch the Pi-GUI demo bounce between computer use and an integrated terminal, inspect the subagent status screenshot down to token counts and artifact paths, grab a voice layer for Pi, and start from templates instead of scratch with the extension starter repost and a browser-agent extension build repost.

Pi-GUI's new beta is the cleanest sign that Pi is turning into a UI substrate, not just a terminal harness. mattlam_'s post says v0.1.0-beta.26 adds both computer use and an integrated terminal, then uses an extension compatibility patch as the showcase for how quickly those pieces can be wired together.

The interesting detail is the workflow, not the feature label. The demo flow goes: find an OSS extension on X, ask an agent to audit it for security, then ask the same stack to patch Pi-GUI so the feature works there too, according to mattlam_'s description.

pi-subagents shipped the most concrete new operator primitive in the batch. According to nicopreme's weekend ship note, the package now bundles /parallel-review for instant parallel reviews and fixes earlier interruption issues between pi-intercom and subagents.

The attached terminal capture adds the mechanics that the tweet text only hints at:

• one run fans out into three roles: two reviewers and one verbosity cleaner, in nicopreme's screenshot
• each agent reports its own tool uses, token count, runtime, and last-active state, in the same screenshot
• each branch writes a separate markdown artifact under a subagent-artifacts path, in the same screenshot

The text of nicopreme's post adds the control surface around that view:

• /subagents-status for run inspection
• /agents updated with fork, background, and worktree support
• parallel tasks can now read files, save outputs, and keep progress notes

A lot of the weekend shipping was scaffolding. badlogicgames's repost of the extension starter points to an end-to-end type-safe starter for Pi-powered Chrome and Chromium extensions, while badlogicgames's repost of a Docker workers build describes an app that spins up Pi workers in Docker by default.

That sits next to badlogicgames's repost of 0xSero's agent-loop claim, which argues that pi-mono/agent is small enough to teach from because the loop lives in only a few files. The shared pattern across the posts is modularity: UI extensions, containerized workers, and a compact agent loop instead of one giant monolith.

Pi is also being used as a local orchestration layer, not only a hosted-model shell. badlogicgames's repost of a DeepSeek-V4-Flash run claims four parallel agents on Pi running on an M3 Ultra at roughly 30 to 34 tokens per second and 160 to 180 watts.

That performance post lands next to the new subagent controls in nicopreme's pi-subagents note, which makes the weekend's throughline pretty concrete: people are pairing Pi's agent loop with parallel execution, local inference, and better inspection instead of treating it as a single-agent chat shell.

The smallest posts are still useful because they show where the add-on surface is heading. nicopreme's TTS link post points to voiPi as a free text-to-speech option for Pi, and badlogicgames's repost of a browser-agent build describes telling Pi to install agent-browser and build its own extension.

Taken together, those posts add two more pieces that do not show up in the core harness itself: voice output and browser-native assistant behavior. That makes the ecosystem story broader than one feature ship, because the new pieces span desktop UI, review workflows, extension scaffolds, local runtimes, and peripheral add-ons.