Skip to main content
Probe Research records the work, not just the numbers. A training run and its loss curve, yes — but also the eval sweep, the dataset pass, the literature review, the design decision someone made on a Tuesday, and the coding-agent session where all of it actually happened.

​
Start here

One command sets Probe up on this machine. It signs you in, installs what your coding agents need, and offers to import the work you have already done.

Walk through the setup wizard

Four screens, about two minutes. Then everything below is available to you.

Sign in

The wizard opens your browser and you approve the device there. Nothing to paste.

Pick your coding agents

Claude Code, Codex, pi, Kimi Code, or several. Detected ones are pre-ticked. Capy runs in the cloud, so it has its own setup.

Review and install

Exactly what will be set up, in plain terms, before anything happens.

Import what you already have

Saved coding sessions, a folder of project work, or both. Nothing uploads without your say-so.
Already working in a coding agent? It can sign you up or in itself: it follows useprobe.ai/auth.md, and your only browser step is signing in with Google and typing the 6-digit code the agent shows you. See agent sign-in.
Only instrumenting a training script, with no coding agent involved? pip install probe-research and go straight to the SDK quickstart.

​
Three ways in, one backend

Probe has three write surfaces. They are not alternatives to each other — most teams use all three, because they cover different moments.

Agent plugin

From a coding agent. Skills that teach Claude Code, Codex, pi or Kimi Code the workflow, plus a read-only MCP server so the agent can query lab history.

Python SDK

From inside running code. Metrics and spans as a training or eval script executes. The replacement for a wandb.init() call.

CLI

From a shell. Register projects and runs, attach files, write notes, import history, read anything back.
Everything lands in the same place: the dashboard at research.prbe.ai, where the work is browsable, chartable, and searchable.

​
What Probe holds

Projects, experiments, runs and run groups; metric series with arbitrary coordinate axes; spans and trajectories; immutable experiment versions; lineage edges between runs, artifacts and papers.See projects and experiments and metrics.
Artifacts anchored to a run, experiment or project — content-addressed, deduplicated by SHA-256, versioned. A data catalog of datasets, sources and checkpoints, with where each came from. Plus non-disruptive code and environment snapshots that can be restored later.See artifacts, datasets and snapshots.
Three kinds, deliberately separate: a short description for identity, a teammate-facing summary page that the recorders write as work happens, and agent-facing notes with titled sub-notes and merge-on-push editing. Plus a shared team note injected into every agent session.See summaries and notes.
With capture on, your Claude Code, Codex, pi and Kimi Code sessions stream into the team knowledgebase: prompts, replies and the shell commands you ran. Your Capy threads come in every 10 minutes once you run probe capy connect. They become searchable evidence for why something was done.See session capture.
Papers recorded against a research project — what each says, the repo read alongside it, and where the two disagree — and the links between them, drawn as a tree of how the reading went.See papers.

​
Search that reads the whole record

One index covers all of it. An exact channel matches identifiers, names and paths; a semantic channel matches meaning. Both answer the same query, sectioned by source, with per-result provenance. That is what makes the rest worth doing: a run you recorded in March is findable in September by describing it, and so is the session where you decided to run it. In the dashboard, Search finds work by name and Ask finds it by meaning.

Search and knowledge

Search by name, Ask by meaning, the filter grammar, and what the index actually covers.

​
Where to go next

Core concepts

Workspaces, projects, experiments, runs — and how they nest.

Record a training run

Three lines: probe.init, probe.log, probe.finish.

Import from Weights & Biases

Local run directories or a hosted project.

Run it on our own cluster

Self-host from the same Helm chart.