Kalmia finds errors hiding in your agent traces
We triage your traces and spawn specialized investigator agents to catch and contextualize your hidden agent errors.
“For auditing agent behavior, Kalmia is now the best tool we have, and it isn't close.”
“Using Kalmia, we've been able to cut total API spend in half.”
Aayush · Founder & CEOHow Kalmia catches agent errors
Detector agents triage your traces
Detector agents triage your traces and cluster what looks wrong.
Investigator agents dig in
Investigator agents dig through the trace and find errors other tools miss.
Kalmia learns your system over time
Findings become knowledge that improves future investigations.
Connects to the
tools you already use
Keep the observability stack you already run: whatever emits your traces today can send them to Kalmia too.
Agent failures and needs-attention findings reach you as alerts and digests in Slack, Microsoft Teams, or email, with the evidence attached.
Kalmia is MCP-native.
Claude Code, Cursor, or any MCP client can pull Kalmia's insights as context, so your coding agent fixes errors and hill-climbs your agent.
Ask, don't dig. The same read tools our agents use, in your editor.
curl https://kalmia.dev/api/traces?verdict=failed \ -H "Authorization: Bearer $KALMIA_API_KEY" { "traces": [{ "id": "22c32b3c", "verdict": "failed", "root_cause": { "span": "#1102", "title": "collision check ignored pending requests" }, "citations": ["#920", "#1014", "#1094"] }] }
Everything is an endpoint. Anything the UI can do, an API key can do too.
More tools for agent debugging
Live run watching Built for long-horizon agents: watch behaviors and errors as the run goes, not after it ends.
Prompt audit Find where your system prompt and your traces disagree.
Behaviors and rules Describe a rule in plain language. Runs that break it fail.
Cost and latency Cost and latency for every run, and totals per agent version.
Variant comparison Run the same workload across prompts and models, and compare.
See what's hiding in your traces
Instrument in two minutes with our SDK or your existing observability stack.
Book a demo
Langfuse