telemetry.dev: LLM observability for AI agents
On October 31, 2026, telemetry.dev launched on Product Hunt under the Analytics, Developer Tools and Artificial Intelligence tags, collecting 235 upvotes and 22 comments. The hosted platform is built on OpenTelemetry. Each agent execution is grouped into one trace, and the individual model requests and tool invocations inside it appear as spans. Each span reports its runtime, token usage, spend and any failure, so an engineer can pinpoint the failing step and the cost of a specific run.
The launch adds three capabilities beyond tracing. Prompt management lets teams change prompt versions without shipping new code. Evals run a candidate prompt against a dataset and compare its scores with production traces, so testing and live monitoring share one data store. Alerts flag error bursts, cost spikes and agent loops. A command line tool and a read-only MCP server let a coding agent query the traces directly. Teams connect through TypeScript or Python SDKs or an existing OpenTelemetry exporter, with listed support for OpenAI, Anthropic, the Vercel AI SDK and LangChain.
The free plan costs nothing and requires no credit card. It allows 30,000 ingestion units a month, 14 days of retention, one project and two seats, with one unit counted per span, log record or metric data point. The Python package reached PyPI on July 31, 2026, and its latest release, 0.2.9, followed on October 14. The Product Hunt listing names Ephraim Atta-Duncan as one of the makers.
Why it matters
Agent costs and failures are hard to attribute without step-level data. A team that cannot see which model call consumed tokens or which tool loop repeated cannot price its product or protect margins with confidence. Cost and loop alerts give operators a way to cap spend before an invoice or a customer complaint reveals the problem. For founders, the link between prompt changes and evals means a change can be scored against production-like data before it reaches users, which can shorten release cycles for small teams.
For investors and buyers, the category is crowded and a new entrant has to show retention and paid conversion, not launch-day attention. The 235 upvotes signal interest, not adoption. Teams evaluating tools should ask whether the vendor publishes overhead benchmarks and data location details, since a tracing layer sits on every request and every byte of prompt data passes through it.
Key points
- 1Each agent run is recorded as one trace, with every model call and tool call shown as a span.
- 2Each span carries duration, token counts, cost and error status for that step.
- 3The free plan is $0 with no credit card, 30,000 ingestion units monthly and 14-day retention.
- 4Prompt versions can change without a deploy, and evals compare dataset scores with production.
- 5
