obsrv
Features

Everything obsrv ships, in one page.

The reliability and observability surface for AI runtimes — built for agents, multimodal systems, and computer-use workflows that need to ship to production.

Trace & replay

  • Step-tree timeline
    Auto-parented steps with model, latency, tokens, and cost — collapsible like a stack trace.
  • Live tail
    Watch traces stream in over Server-Sent Events as they ingest.
  • Replay viewer
    Reconstruct full sessions including computer-use desktop and browser flows.
  • Trace filters
    Status, run type, model, tag, and arbitrary metadata JSON paths — all server-filtered.

Evaluations

  • Synthetic metrics
    LLM-as-judge metrics for adherence, coherence, tool selection, and unsupported request handling.
  • Observed signals
    Capture refunds, escalations, thumbs-down, and any custom event you record.
  • Per-release scoring
    Tag traces by release and prompt version. Compare regressions in one click.
  • Annotations & flags
    Mark runs for review, attach notes, and roll them into evals.

Cluster discovery

  • Continuous embedding
    Every trace embedded as it lands. Clusters update with traffic shifts.
  • No predefined categories
    Patterns surface from real behavior — not a template guess.
  • Auto-labelled
    Cluster names generated from representative behaviors via Claude.
  • Drill to traces
    Jump from cluster to underlying runs and replay them in context.

Multimodal & computer-use

  • Native rendering
    Image, audio, video, sensor, and file artifacts inline in the timeline.
  • Browser/desktop replay
    Reconstruct sessions from captured screenshots and action streams.
  • Tenant-scoped storage
    Object keys follow orgs/{org}/projects/{project}/traces/{trace}.
  • Browser-safe media
    Stream artifacts through signed proxy URLs — never expose raw object keys.

Monitors & alerts

  • Threshold monitors
    Watch metric pass rates, latency, error rates, and cluster volumes.
  • Webhook delivery
    Push alerts to Slack, PagerDuty, or any endpoint that speaks JSON.
  • State transitions
    Open, acknowledged, snoozed, resolved — with full history.

Developer experience

  • Python & Node SDKs
    Symmetrical APIs, async ingest, fail-soft defaults.
  • Provider integrations
    OpenAI, Anthropic, LangChain, OpenClaw — drop-in wrappers.
  • MCP server
    First-class agent tooling for IDE workflows.
  • OpenTelemetry export
    Pipe traces into your existing OTel collector.

Operations & governance

  • Org & project isolation
    Every API key scoped to a project. Storage paths follow tenancy.
  • API key hashing
    argon2id at rest. SHA-256 fingerprint lookup.
  • Retention controls
    Per-project retention policies for traces and artifacts.
  • Usage attribution
    Per-org and per-project ingest, storage, and query usage.
Read the docs →
08 — DEPLOY

Ship the agent.
obsrv will record everything.

Drop the SDK in three lines. The recorder lights up the moment traffic starts flowing.

Book a callSign in
$pip install theta-obsrv·npm i @theta-lab/obsrv
FDR · OBS-1
REC
tr_01HXR4Z9CK
support_agent · 4.2s
✗ wrong_order