We started obsrv a year ago with a simple bet: that the next decade of software is going to be built by agents, and the teams shipping those agents are going to need observability that takes the runtime seriously — not a tracing dashboard with an LLM badge stapled on.
Today, after twelve months in private beta with thirty design partners, obsrv is generally available.
What you get on day one
- Trace replay for any agent, with auto-parented step trees, multimodal artifacts, and live tail.
- Continuous cluster discovery — failure modes surface from real traffic, not from a category list someone picked in a sprint planning meeting.
- Synthetic + observed evaluations, so the metrics on your dashboard reflect what your users actually felt.
- Computer-use replay for browser, desktop, and mobile sessions — full screenshot timeline, action stream, every frame.
- SDKs in Python and Node, with first-class adapters for OpenAI, Anthropic, LangChain, OpenClaw, and the Vercel AI SDK.
- OpenTelemetry export and an MCP server for IDE-side agent workflows.
What we believe
The trace is the source of truth. Multimodal evidence is a first-class requirement, not an add-on. Tenancy boundaries belong in storage, not just in the dashboard. The SDK surface should be small, predictable, and never the reason your application goes down.
We're going to keep saying these things until they're obvious — and shipping the product that makes them feel obvious.
What ships next
A unified action/event renderer for computer-use traces. A real monitor execution engine with state transitions and suppression. Regression detection tied to releases, prompts, and models. Audit logs, fine-grained RBAC, SSO, and SCIM. Native OTel ingest. BYO storage on Enterprise.
If any of that is the missing piece for your team — come talk to us. If you just want to start instrumenting today, you can be live in three lines:
pip install theta-obsrv
# or
npm i @theta-lab/obsrvWelcome to the runtime.