From endpoint to evidence.
Collect measurements locally, keep the samples, and build a comparable record in your workspace.
Create your workspace
Open the workspace, select Create account, and use a unique password with at least 12 characters. Sign in on the same site to return to your saved records.
Choose Add a run to enter measured latencies, or Import JSON to upload a file produced by the CLI.
Run a local measurement
Requires Node.js with built-in fetch and parseArgs support. Run node benchmark.mjs --help to inspect the exact options.
node benchmark.mjs --endpoint http://localhost:8000/v1 --model YOUR_MODEL --hardware "Your environment" --requests 10 --warmup 1 --out run.json
The endpoint must be a chat-completions-compatible HTTP(S) base URL. Replace YOUR_MODEL and the example endpoint with an authorized workload. API requests can incur your provider’s charges.
If authentication is required, set MODEL_API_KEY in your local shell. The CLI keeps the key local and does not write the prompt or endpoint into its run file.
Import and compare
Import the resulting run.json into the workspace. A run includes a name, model, hardware label, source, notes and samples. Each sample contains latencyMs and optional outputTokens.
Select two equivalent runs with their comparison checkboxes. Differences are descriptive, not a statistical-significance test; use equivalent prompts, settings and environments.
Export and remove
Use Export CSV for summary rows or Account → Export all run data for JSON with raw samples. Individual runs can be deleted from the table.
Account deletion removes the active account, sessions and run records. Export first if you need a copy. See the workspace API reference for the supported HTTP routes.