Welcome to RLViz

RLViz is an open-source viewer for agent rollouts: model output, tool calls, observations, graders, rewards, errors, and source provenance in one local workspace.

Browse reviewed public trajectories at benchmarks.rlviz.dev. The catalog keeps benchmark provenance and quality gaps visible before opening an exact, digest-pinned bundle here.

Please read

The hosted browser directly supports reviewed .rlviz bundles, RLViz canonical NDJSON, Letta trajectory v1 JSON, Harbor ATIF v1.5-v1.7, Inspect AI EvalLog JSON, and Verifiers GenerateOutputs JSON. A selected browser directory must contain exactly one supported .ndjson or .json trace export; open an .rlviz bundle through the file picker, drag it onto the viewer, or follow an explicit-consent public HTTPS link pinned by the complete file SHA-256. Complete Harbor job directories and bundle creation use the native CLI.

We highly recommend installing the local CLI. It opens complete Harbor jobs as one read-only source. Coding agents can operate its workspaces and build adapters or plugins for private formats, larger cohorts, and other multi-file datasets. Browser adapters run as reviewed WASM modules; local adapters can use the source system's native files and libraries.

Install the local CLI

Homebrew:

brew install TheSnakeFang/tap/rlviz

npm:

npm install --global rlviz

Verified shell installer:

curl -fsSL https://rlviz.dev/install.sh | sh

Then configure RLViz and open a trace:

rlviz init
rlviz open ./path/to/trace.ndjson

The CLI adds persistent indexing, larger cohorts, private adapters, structured queries for coding agents, and named workspaces that an agent can update after the GUI opens.

What it does