Quickstart

Start with the wizard. It is the shortest path in, and it is the only step that needs doing once per machine.

Set up with the wizard
Record a training run

Run it

npx probe-research install

No prior install needed — npx fetches the launcher, which resolves the real CLI. The wizard signs you in through the browser, then walks four screens: pick your coding agents, review what will be set up, install, and offer to import work you already have. Before the import offer it asks where this computer’s work should go — the workspace its new projects are filed in.

Add --agent codex to set up Codex instead of (or as well as) Claude Code. --agent both does both. Everything the menu offers is also a flag — see unattended use.

Confirm it took

probe doctor

Read-only. It prints what is installed, which credentials resolve, whether auto-update is working, and when the last update attempt was.In Claude Code, the install also added a one-line indicator under the input box, such as ● tracking, ● tracking → folding or ● tracking → folding · running. See Status line for every state.

Use it

In your agent, the switch is /probe:

/probe on      reads and writes
/probe read    search yes, record no
/probe off     no calls at all

With the switch on, your work is recorded as it happens — by the Probe daemon beside your agent, or by the agent itself through the /track-work skill, depending on the wizard’s Who records setting. The agent can also search your lab’s history read-only at any time through the MCP server.

Bring your history with you

From the wizard’s main menu, choose Import research work — or run it directly:

probe backfill /path/to/project --project existing-project
probe backfill --transcripts-only

Folder imports read the directory, upload what they find and describe it. Session imports pull saved coding-agent transcripts into the knowledgebase. Both run in the background and survive a terminal disconnect.

The wizard in full

Every screen, every menu row, and the flags behind them.

Install the client

pip install probe-research

This gives you both the Python SDK (import probe) and the probe command-line tool. They are the same implementation with two ergonomics.

probe wizard --action login

This opens a browser device flow and signs the device in — nothing else is installed. You approve the device in the dashboard; nothing is pasted. The credential is written to ~/.config/probe/config.json.Signing in this way does not ask which workspace this computer’s work goes into. If your team has more than one workspace, pick one before the first run:

probe workspace list
probe workspace use <slug>

You can skip this step entirely. The first probe.init() or probe run start with no token triggers the same browser approval inline, as long as you are on a TTY. In CI, set PROBE_TOKEN instead — headless callers get a clean AuthError rather than a hung prompt.

Instrument the script

import probe

probe.init(
    project="folding",
    experiment="dockq-sweep",
    question="does temperature 0.7 win?",
)

for step, batch in enumerate(loader):
    loss = train_one(batch)
    probe.log({"train/loss": loss}, step=step)

probe.finish()

project and experiment are created if they do not exist. question is the one opt-in to creating an experiment, so the first run in a new experiment says what you expect to see; every run after that is bare.A new project goes into this computer’s workspace. On a team with several workspaces and none set, probe.init raises WorkspaceRequired before creating anything: run probe workspace use <slug>, set PROBE_WORKSPACE, or pass workspace= to probe.init.If the script exits without finish(), the run is still closed at exit as completed, failed or canceled rather than left for the crash reaper.

Watch it

Open research.prbe.ai — the run appears under its experiment with live curves. Or stay in the terminal:

probe metrics plot $RUN_ID --key train/loss

What you just created

workspace
└── project          folding
    └── experiment   dockq-sweep        ← carries the question
        └── run      run-1              ← carries metrics, files, spans

A run does not need an experiment — a project-direct run is the honest home for work that has no hypothesis. And an experiment does not need to be created up front; naming one in probe.init() with a question creates it.

Core concepts

The full model, including workspaces, run groups, artifacts and notes.

Next steps

Log more than scalars

Coordinates, labels, non-numeric values, derived series and terminal plots.

Capture files and code

Artifacts, anchors, versions and reuse.

The whole CLI

Every command group, with what each one is for.

Import from W&B

Local run directories or hosted projects.

Getting started

Setup wizard

Tracking your work

Python SDK

CLI

Coding agents

Dashboard

Integrations and imports

Deploy and operate

Reference

The wizard in full

What you just created

Core concepts

Next steps

Log more than scalars

Capture files and code

The whole CLI

Import from W&B

The wizard in full

​What you just created

Core concepts

​Next steps

Log more than scalars

Capture files and code

The whole CLI

Import from W&B

What you just created

Next steps