# An agent-native Weights & Biases alternative

When coding agents run most of your experiments, the tracker has to work for them too: set itself up, take their writes, and answer their questions before they start the next run. Probe does that, and it keeps your W&B history.

Tell your coding agent:

```
Install Probe on this machine and its coding agents. Read https://research.prbe.ai/install.md and follow it.
```

## Switch without losing anything

- **Same shape as W&B.** `probe.init()` and `probe.log()` take the place of `wandb.init()` and `wandb.log()` in a training script. See the [SDK quickstart](https://docs.prbe.ai/sdk/quickstart).
- **Full W&B tracking parity.** Anything you track in Weights & Biases, you can track in Probe.
- **Bring your history.** Run `npx probe-research` and pick **Import research work**, or import from the command line:

```
probe wandb import-local /srv/work --project folding
probe wandb import-hosted my-team my-wandb-project --project folding
```

See [Weights & Biases in the docs](https://docs.prbe.ai/integrations/wandb).

## What you can do that a dashboard can't

| You want to | With Probe |
| --- | --- |
| Compare runs without building a dashboard | Ask the dashboard agent "what changed from the baseline?" and get a chart linked to its source runs |
| Let your coding agent read past results | `probe ask` from the session, or a hosted MCP server with six read-only tools for any MCP client |
| Stop agents repeating failed experiments | Agents read the team's projects, experiments and notes before they start |
| Track RL, not only scalars | Rewards per step, plus rollouts, turns and tool calls as spans, and Harbor trials |
| Keep the record after a crash | Data already saved stays available after the run or the machine stops |
| Get set up without a setup guide | Your agent signs you up and installs Probe; you sign in once in the browser |

## Built for teams running agents

- **One record per team.** Projects, experiments, runs, notes and the coding-agent sessions behind them live in one place your team shares.
- **Your agents start informed.** Each session can read what teammates and other agents already tried.
- **You stay in control.** `/probe on`, `/probe read` and `/probe off` decide per conversation whether Probe records.

## Learn more

- [Experiment tracking for Claude Code, Codex and Cursor](https://www.useprobe.ai/experiment-tracking-for-coding-agents)
- [RL experiment tracking for rollouts and rewards](https://www.useprobe.ai/rl-experiment-tracking)
- [Probe vs MLflow](https://www.useprobe.ai/compare/mlflow)
- [Probe vs Trackio](https://www.useprobe.ai/compare/trackio)

Source: https://www.useprobe.ai/wandb-alternative
