Orchestra

Your own AI team, running privately on your computer.

You ask for something in plain words. A composer plans it, workers do the jobs, and the results land in your work folder as files, PDFs and pictures. It all runs locally on Ollama models picked to fit your PC, with OpenClaw agents. Your prompts and your work never leave your machine; the only downloads are the models themselves, during setup.

How it works

You ask. Simple jobs ("write…", "summarize…", "translate…", "draw…") go straight to a worker.
The composer plans. Bigger asks are split into tasks and assigned across the team.
Workers deliver. Each task names the file it has to produce, and it only counts as done once that file exists.

What it does

A team sized to your PC. Big graphics cards get big models. Laptops and PCs without one get small ones, with a single model for every job so nothing swaps in and out of memory (Light mode).
Files, PDFs and pictures in your work folder, with a history you can restore from.
Memory and playbooks: notes every agent reads, facts it learns from your conversations, and step-by-step playbooks for jobs you want done the same way every time.
A live dashboard showing what each agent is doing, plus tasks, metrics, storage, models, tools and settings.
Guardrails in code, not prompts. Small models get firm limits: no duplicate tasks, no placeholder jobs, and nothing marked done without the file to prove it.

Get it

Your computer	File
Windows 10 / 11	`Orchestra-Setup-x.y.z.exe`
Mac with Apple silicon (M1 and later)	`Orchestra-x.y.z-arm64.dmg`
Mac with an Intel chip	`Orchestra-x.y.z.dmg`
Linux	`Orchestra-x.y.z.AppImage`

On first start Orchestra checks what it needs (Node.js, Ollama, OpenClaw), offers to install anything missing, scans your hardware and downloads models that fit. Windows and Linux copies update themselves; on a Mac, grab new versions from Releases.

First-run warnings: the installers aren't code-signed yet. On Windows, click More info, then Run anyway. On macOS, right-click the app, then Open.

What it needs

Windows 10 or 11, macOS, or Linux
About 4 GB of free memory at minimum; 8 GB or more works much better. A graphics card helps but isn't required.
3–20 GB of disk space for the models
An internet connection for the first setup; after that it works offline

Under the hood

One Node.js/TypeScript service with an Electron desktop app. Tasks and the team board live in a local SQLite store, agents talk to the service through an orchestra-team OpenClaw plugin, and the dashboard runs on your own machine. It's open source under the MIT license; build notes are in the README.