Launched this week
Oodle AI is observability built for agents: one platform for agent observability, logs, metrics, traces, APM, RUM, and infrastructure. Oodle gives agents full production context to debug apps, infra, and other agents. Built on open standards. Object storage for telemetry. Serverless compute for queries. Sub‑second agentic queries at 80% lower cost. Trusted by LanceDB, Truefoundry, Fello, WisdomAI, CultFit and more. 3PB+ of telemetry and 900M+ agent spans/month, zero sampling.








Free Options
Launch Team

Transformer.gg by Jon Oringer(founder of Shutterstock)The agentic influencer marketplace for brands, influencers, and bots.
Promoted

Hello Product Hunt folks!
We’re the team behind Oodle, and today we’re launching Oodle Agent Observability.
Debugging AI agents is hard. Your agent could be stuck in a loop, wasting money on model calls, or making decisions that hurt your customers. Now imagine debugging millions of agent runs in production.
You need to know what each agent did, where things went wrong, and what happened across your applications and infrastructure, without spending a fortune on observability.
We believe every agent run matters. That’s why we built Oodle on DenaliDB, our own serverless, object-storage-native database engine, designed to capture every trace without sampling at any scale.
Here’s what makes Oodle different:
Debugging at scale with full context: See what every agent is doing and investigate issues across agents, applications, and infrastructure in one place.
3–5x lower observability costs: Keep every agent trace without sampling, powered by DenaliDB's fundamentally different architecture.
Hierarchical evals: Evaluate individual steps, tool calls, and final answers to pinpoint where agents go wrong.
Datasets and experiments: Use real agent runs to test changes, compare results, and improve performance.
Proactive insights out of the box: Automatically analyze traces for sentiment, classify issues, and detect anomalies. Oodle’s AI agent verifies potential problems, filters out false alarms, and highlights issues affecting your customers.
Our mission: Help teams build AI agents they can trust in production, without sacrificing visibility or breaking the bank.
Last quarter, Oodle processed 9 petabytes of telemetry and 1.8 billion agent spans across AI-native startups and public companies, including LanceDB, Truefoundry, WisdomAI, CureFit, Fello, Vorflux, Fuel, and Lookout.
And we're just getting started!
If you're building AI agents, we'd love for you to try Oodle and tell us what you think!
- Team Oodle
We've been using Oodle at AmpUp for almost a year now. I just paste in a bug: this broke, go figure out what went wrong. Oodle figures it out most of the time!
@rahul_goel5 Agents with right context provided to them can do wonders. Great to be part of your journey, Rahul.
What is DenaliDB? Is it open source?
Also query latency at scale?
@arvi_ Thanks for your interest.
DenaliDB is our in-house built database engine for Observability. It is built on top of object storage and has a serverless query engine.
It is not open source, but we have a number of talks on it:
1. @tao_wan sharing the architecture and learnings in Apache Data Fusion meetup: https://youtu.be/llOtoL4b-Bg?si=etvCN9MwGclkOVO-
2. @sujeet_kausallya_gholap talking about the design in HasGeek conference:
On query latency at scale, @mvijaykarthik recently also benchmarked DenaliDB against ClickHouse for metrics workload, here is a blog with the setup and results: https://www.oodle.ai/blog/has-clickhouse-conquered-metrics .
@arvi_
DenaliDB supports metrics, logs, distributed traces and agent traces. It's an all-in-one database for observability.
Architecture blog: https://www.oodle.ai/blog/how-oodle-keeps-observability-fast-at-scale
Decoupling storage from compute is the right way to bend observability cost curves especially in today’s workd - curious to see how it holds up on high-cardinality workloads. Nice work!
@saraswathy_hariharakrishnan Thank you! Regarding high cardinality, If you think of a trace as a set of label-value pairs, it's high cardinality by nature, session IDs, user IDs, prompts, tool calls. Our custom columnar format reads only the columns a query needs, so searches across all your agent traces stay quick and efficient, even at scale. This is also what allows us to efficiently store and manage Agent sessions which are GBs in size.
Do I pay for LLM-as-judge eval tokens?
@pruthvi You can bring your own API Keys for LLM-as-a-judge evals.
To reduce LLM eval costs, you can also chain your LLM evals behind code evals. One of our customers Fello reduced their LLM eval costs by 80-85% using hierarchical evals: https://www.oodle.ai/case-studies/fello-agent-observability
Congrats @gaurav_maheshwari8 and @sujeet_kausallya_gholap . Not needing to sample is great but how does the MCP integration work with huge volumes of data for analysis and insights? Wouldn't it flood the LLM context?
@tirumarai Great question.
Sampling and Context available to agents for debugging are independent dimensions. Because at petabytes scale, even sampled data would be much higher than what AI agents can process. Unlike APM use-cases, where tail sampling is a good way to throw away traces which are not that significant (no errors, low latency etc.), there isn't any good way to know whether an agent trace is significant or not.
Regarding context for agents to debug, we solve it in few ways:
1. Proactive insights on signals like user dissatisfaction, tool call failures, multiple turns etc. Oodle MCP surfaces these insights to the agents where agents can see insights for a particular agent, or service, or a particular signal category: https://play.oodle.ai/genai/traces/?tab=insights
2. Chained evals: Oodle supports code evaluators and LLM-as-a-judge evaluators and a pipeline can be constructed to run through traces with code evaluators and only pass down a filtered set of traces (based on code eval scores) to expensive LLM-as-a-judge.
https://www.oodle.ai/blog/you-cant-sample-your-way-to-reliable-agents covers it in more details.
Congratulations, Kiran, Gaurav and the Oodle team!
Having worked on the infrastructure and agentic side of things ourselves, I know how difficult getting meaningful production context can be.
Loved seeing what you guys have been building, and even more excited to see it out in the wild.
Rooting for you guys! 🚀
@sir_priyank Thanks Priyank for your support in our journey!
@sir_priyank Thank you! Appreciate kind words.
Oodle MCP has been very useful for us, to debug production logs via claude easily.
@harshit_kedia1 Thanks Harshit, great to hear you have been finding Oodle helpful.