You add an agent to a channel the same way you add a person. Each agent has its own keys (its own identity), its own channel memberships, and its own audit trail (a record of everything it did).

Add the agents you already use (buzz-cli hooks into Goose, Codex, and Claude Code today) or create new ones right in Buzz — each with its own identity, channel memberships, and audit trail.

Hook up a repo and a feature branch becomes a channel — code patches, test results, an agent's first-pass review, and the merge decision all land in the same thread.

Agents run in Spaces — parallel workspaces inside the same browser — so Claude Code can work 10 tasks in 10 Spaces in the background while you keep browsing up front. You can watch, take over, or stop any Space.

On first launch it offers to migrate your Chrome data: logins, cookies, extensions, bookmarks. Say yes and the agent never hits a login wall on sites you already use.

Browser actions are exposed as JavaScript functions the agent calls in one pass, instead of a slow command-look-command loop. The project's own benchmark claims complex tasks finish up to 2.5× faster with fewer tokens (the words AI models are billed by).

Per the README, it ran for two years as Alibaba's internal review assistant across tens of thousands of developers and millions of caught defects — then was “incubated” into open source.

Fixed engineering logic (deterministic pipelines) handles the steps that must be exact — picking which files to review and pinning comments to the right line — while the LLM agent (the AI reasoning part) handles the actual judgment calls.

On its own benchmark — 200 real pull requests from 50 popular open-source repos, checked by 80+ senior engineers — it beats a general-purpose agent on precision using about 1/9 the tokens (the units AI usage is billed in). It deliberately trades away some recall: it flags less, but what it flags is real.

Lead with the next action, number the steps, cap lists at five items, give time estimates in minutes, end with one concrete next step — and no preamble, no recap, no “Hope this helps!” closers.

It ships as a Claude Code plugin (two claude commands, then type /i-have-adhd) and a Codex plugin; no local download needed, and Claude Code keeps it updated. Instructions for other agents live in INSTALL.md.

Loosely based on The Adult ADHD Tool Kit by J. Russell Ramsay and Anthony L. Rostain — adapted for how an AI should answer, not how a human should organize their day.

Tools for walls, floor slabs, ceilings, roofs, zones, and items like doors, windows, and lights — organized as Site → Building → Level, with stacked or exploded floor views.

Built with React Three Fiber and WebGPU (the browser's fastest new graphics engine), with a live editor hosted at editor.pascal.app.

Scenes persist locally in the browser (IndexedDB, the browser's built-in storage) with 50 steps of undo/redo.

Run /book-to-skill on a PDF, a folder, or a list of files and it distills the content into a core skill file, one file per chapter, a glossary, a patterns list, and a cheatsheet — structure, not a summary.

Ask about “replication” and the agent reads just that chapter (≈1,000 tokens) instead of re-paying for the whole 200,000-token book every message.

PDF, EPUB, DOCX, Markdown, HTML, RTF, and MOBI are supported — the README pitches internal docs, brand guidelines, research-paper stacks, and specs as inputs too.

27.8 MB of RAM per session and a 14 ms start, per the README's own benchmarks — the pitch is running many agent sessions at once without your laptop melting.

Every turn is embedded as a semantic vector (a math fingerprint of meaning), so relevant past context is recalled automatically — no memory commands, no token burn.

Run several agents in one repo; a server coordinates them, warns an agent when another edits a file it read, and agents can even spawn their own teammates.

The flagship /grill-me skill makes the agent relentlessly interview YOU about what you want before it writes anything — Matt calls these his most popular skills, built to fix the #1 failure mode: the agent building the wrong thing.

/grill-with-docs also builds a project glossary (a CONTEXT.md file) so the agent stops using 20 words where 1 will do. The README calls it possibly “the single coolest technique in this repo.”

A /tdd skill enforces red-green-refactor (write a failing test first, then make it pass), and /diagnosing-bugs wraps hard debugging into a disciplined reproduce-to-fix loop.

It doesn't sell AI — it controls Claude Code, Codex, Cursor, Grok Build, and OpenCode if they're set up and logged in on your computer.

Ships a mobile app for iOS and Android (the README calls it best-in-class), plus a web app and an Electron-based desktop app, so you can steer agents on your machine from anywhere.

The README answers its own question: nothing. It's MIT-licensed and deliberately open so that if the project ever goes the wrong way, you have everything you need to fork it.

Curated feeds across 15 categories get synthesized into briefs by AI — and it can run entirely on local AI via Ollama (models on your own machine), no API keys required.

Dual map engine: a 3D globe and a WebGL flat map (browser graphics tech) with 56 map layer types, plus live flight data provided by Wingbits.

Its own server-computed stress score (CII v8) for 31 Tier-1 countries, alongside a finance radar covering 29 stock exchanges, commodities, and crypto.