Your AI agent writes the code. BuzzAgent shows you every diff, every command, every step β approve or veto in one click. The strongest open agent core, wrapped in the interface it deserves. β‘
Windows Β· macOS Β· Linux β one installer, the agent core included. π¦
The problem π§
The market splits in two: visual-but-weak assistants locked inside IDEs, and powerful-but-blind terminal agents. BuzzAgent refuses the trade-off β a full visual workbench over the strongest open agent core, running on your machine, calling no server but yours and your provider's. π
See every move. Approve every step. π¬
Streaming text, tool cards with real arguments, live diffs, console errors surfaced as text the agent can act on. Nothing happens on your machine without a visible trace.
What you get β¨
The core does the heavy lifting; the workbench makes it observable, safe and fast.
Tokens land incrementally with markdown and syntax highlighting β no reflow, no flicker, 60 fps in thousand-message sessions. Inspect every tool call: the exact arguments, the duration, the error if it failed. Never a swallowed spinner.
Real side-by-side git diffs with word-level marks. Review while the agent works β and roll any file back with one click, straight from the panel.
Managed headless Chrome β or attach to your real browser over CDP, logged-in profile included. Screenshots and console stream straight into the agent's context, so it fixes what it actually sees.
Run three agents on three branches at once β isolated git worktrees, one visual monitor. Ship the best, discard the rest. π§ͺ
Shell commands, file writes, browser control β all wait for your explicit yes. Allow once, always for this tool, or deny. You are the gatekeeper. π¦
Anthropic, OpenAI, Google, local Ollama β any OpenAI-compatible endpoint. Sign in with Claude Pro or Copilot via OAuth; keys never leave the core.
Toggle MCP servers per project, build custom agent modes, run /compact, /undo, /export from the composer.
Dictate prompts through your local Whisper server, queue tasks while the agent works, get desktop + ntfy pings when a run finishes.
Early bees π
"Finally, an agent UI that shows everything and hides nothing. The permission inbox alone is worth it."
"I run three worktree agents overnight and review the diffs over coffee. That used to be a full workday of babysitting."
"Zero telemetry was the reason I tried it. The browser-in-the-loop is the reason I stayed."
How it works π οΈ
The app starts a supervised agent core bound to your directory β random port, generated password, health-checked. Your files never leave your machine.
Paste an API key or sign in via OAuth β 75+ models, local Ollama included. Auto-discovery probes the endpoint and confirms it works before you send a thing.
Watch streaming text, tool cards and diffs arrive live. Approve each step β or queue the next task and let it run.
Honest comparison βοΈ
| Capability | π BuzzAgent | IDE assistants Cline Β· Kilo Code |
Terminal agents OpenCode Β· Claude Code |
|---|---|---|---|
| Full visual GUI π» | β | β | β |
| Terminal-grade agent core π§ | β | β | β |
| Zero telemetry π | β verified | β | varies |
| Visual diffs & revert π | β | β | β |
| Browser in the agent loop π | β | β | β |
| Git worktree orchestration πΏ | β | β | manual |
| 75+ providers, OAuth π | β | limited | β |
| Price π | Free, MIT | freemium | varies |
Comparison reflects public product information as of 2026. IDE assistants and terminal agents evolve β so do we. π€
FAQ β
Yes β and verifiably so. Sharing and auto-update channels are force-disabled in config, the audit is public (docs/telemetry-audit.md), and the only network traffic is your local core and the provider you configure.
75+ out of the box: Anthropic, OpenAI, Google, xAI, OpenRouter, local Ollama, and any OpenAI-compatible endpoint. OAuth login for Claude Pro / Copilot-style accounts, plain API keys for everything else.
No. The desktop app ships a supervisor that downloads the pinned core, runs it on a random loopback port with a generated password, health-checks it, and tears it down on exit. One installer, zero terminal.
In the core's own credential store (auth.json), written by the core itself. The UI never stores or even sees a second copy.
Windows 10/11, macOS 12+, and Linux (AppImage / .deb / .rpm). Apple Silicon and Intel both supported.
They are IDE extensions with their own lighter agents and telemetry. BuzzAgent is a standalone desktop app over the full-strength open core β no IDE required, no account, no telemetry, and native extras like worktrees and the browser loop.
Nothing, ever. MIT licensed β free for personal and commercial use, forever. You pay only your LLM provider directly, at their rates, with no middleman markup.
Download β¬οΈ
No account. No credit card. No telemetry. Uninstall is one drag to the trash β that's the whole exit cost.
π οΈ Prefer to build from source?
git clone https://github.com/buzzband/buzzagent.git
cd buzzagent && npm install && npm run setup:core
npm run dev # π desktop app + supervised core
One download, one provider key, one task. If it doesn't click, you've lost sixty seconds and nothing else.
Bugs, ideas, partnership β everything lands in a human's inbox. No trackers in this form either. ποΈ
π Free, MIT, zero telemetry β ready in sixty seconds.
Download free β¬οΈ