w3bs.AccountStart building
W3BS TRAINING · COMMUNITY DRAFT

Teach agents
by doing it yourself.

Record a task on your own desktop, review what was captured, and let only the agents you choose learn from the approved examples. Local first: your recordings stay encrypted on your computer unless you push them to your w3bs.org workspace.

MACOS AND LINUX
curl -fsSL https://w3bs.org/training/install.sh | sh
WINDOWS (POWERSHELL)
irm https://w3bs.org/training/install.ps1 | iex

One self-contained binary, checked against the release checksums. With the w3bs CLI: w3bs install w3bs-training. On Linux, also install libsecret's secret-tool (sudo apt install libsecret-tools) so the encryption key lives in your desktop keyring. Recording needs the Agentic Desktop Layer.

Open your workspace How it works
HOW IT WORKS

Four steps.

  1. Record

    In the Agentic Desktop Layer tray, choose Start training…, say what you are about to do and tick the apps it may see. It keeps pictures of those apps' focused window only. No keystrokes, typed text, clipboard or audio.

  2. Review

    Each recording becomes a task card: confirm the goal, say how it went (complete, partial, failed), clear privacy warnings, split or merge tasks. Nothing is used until you review it.

  3. Share with an agent

    On the Consumers page, Share all reviewed tasks with one named agent (or w3bs training share). It retrieves examples over MCP, as evidence, never as instructions. Private retrieval does not approve model training.

  4. See if it helped

    Agents report each attempt with and without the examples. Failures come back to you, and the Runs page compares the two with confidence intervals instead of claims.

QUICK START

Use it from Claude Code.

Record on your desktop; Claude Code can run on another machine and reads the shared tasks over ssh. Recordings never leave the desktop.

On the desktop

w3bs update                      # or, first time: curl -fsSL https://w3bs.org/install.sh | sh
                                 #   then: w3bs install agentic-desktop-layer w3bs-training
sudo apt install libsecret-tools # Linux, once: the keyring client

Then record (tray → Start training… → Finish task), review in w3bs training open, and click Consumers → Share all reviewed tasks.

On the agent machine

w3bs update
claude mcp add --scope user w3bs-training -- \
  ssh -o BatchMode=yes you@desktop '~/.local/bin/w3bs-training' mcp serve --stdio --consumer claude-code
claude mcp list                  # w3bs-training: Connected

Then ask Claude to use w3bs-training to find how you did the task. Full quick start.

WHERE YOUR DATA LIVES

Two places, one format.

On your computer: w3bs training open runs the review app locally, from a one-time link; frames are AES-256-GCM encrypted with a key from your keyring. On w3bs.org: your workspace is the same app, signed in with your W3BS account, in its own database schema. Send reviewed tasks there with w3bs training push to review on another device or to give a hosted agent access. The server can decrypt what you push to it; keep anything you would not hand to a server on your own machine.

Both speak the open W3BS Demonstration Dataset Format 0.1, so you can export a bundle, a JSONL dataset or next-action training examples (with no future state in the input) and take them anywhere.

Status. Recording is verified on KDE Plasma; GNOME, Windows and macOS use the same code but have not been run through the rig. Recordings are screenshots of approved windows (the visual-review profile), so they help an agent see how a task is done but do not qualify as action-level training data. No model-training integration is offered yet. Read the docs and privacy guide.