Web UI¶

agentlens ui serves a small read-only web UI over the database on http://127.0.0.1:8765/ (and opens it in your browser; --no-open, --port, --host, --db):

Traces: filter by name, experiment, case, metadata (key=value) or errors only, page through them; each row shows pass/fail (with the failed evaluators), latency, cost, steps, tokens and the output. A trace opens on its span tree (the HTML report's viewer).
Experiments: every run of every suite, the baseline marked; pick a baseline and a candidate (or vs baseline) for the comparison: metrics with confidence intervals, gates (agentlens ui --gates gates.yaml) and every case side by side. An experiment opens on its report.
JSON API with the same data: /api/traces?experiment=pr-123&errors=1&limit=20, /api/traces/ID, /api/experiments, /api/experiments/ID, /api/compare?baseline=&candidate=.

agentlens run examples/support_agent/suite.yaml --name main --set-baseline
SUPPORT_PROMPT_VERSION=v2 agentlens run examples/support_agent/suite.yaml --name prompt-v2
agentlens ui --gates examples/support_agent/gates.yaml

Web UI: the candidate's traces, a regressed run flagged Web UI: the regressed run's span tree, the refund call over the limit selected

It runs on the standard library's HTTP server (no extra dependency, no JavaScript framework, nothing loaded from the network), listens on loopback only unless you pass --host, answers only GET, and escapes everything it shows. From Python: agentlens.ui.serve(agentlens.ui.make_server(agentlens.ui.UIApp("runs.db"), port=0)).

API: agentlens.ui.