News

    Introducing Catalyst: Train self-improving AI models

    Learn more

    Trace every step
    your agents take.

    Capture LLM calls, tool calls, framework steps, and your own work as a single trace. Then let Halo, our open-source agent-loop optimizer, read those traces and tell you exactly what to fix. Install in minutes with our AI-powered CLI.

    Talk to an Engineer

    Trusted by the world's best engineering teams.

    Gravity
    Cal AI
    NVIDIA
    24Labs
    Grass
    Rizz

    See exactly what
    your agents are doing.

    Catalyst Tracing captures the full execution of your AI apps and agents (LLM calls, tool calls, framework steps, and any custom spans you add). Drop the SDK into your app, point it at Catalyst, and traces start flowing.

    Drop-in tracing SDK

    Install the SDK, call setup(), and spans start flowing. LLM calls, tool calls, framework steps, and custom work are all captured as OpenTelemetry traces with zero changes to your business logic.

    Works with the frameworks you use

    OpenAI, Anthropic, LangChain, LangGraph, the Vercel AI SDK, OpenAI Agents, LiveKit Agents, Claude Agent SDK, Pydantic AI, and more. We support all the popular agent frameworks and are constantly adding new ones.

    Install with AI

    Use our CLI to instrument your codebase automatically. A coding agent scans your project, wires up the SDK, and gets traces flowing. Review every change before it lands.

    “Our custom model is more accurate, more affordable, and cut request latency by more than 50%. The whole experience was a breeze, and the inference.net team was great to work with.”
    Henry Langmack
    Henry Langmack
    Co-founder, CTO @ Cal AI
    Analysis

    From raw spans
    to what to fix next.

    Browse every trace, group runs by agent, and let Halo decompose your traces into ranked findings with concrete fixes, citing the exact trace IDs that produced them.

    Trace tree view

    Walk the full span tree for any run. Inputs, outputs, messages, tool calls, model metadata, token usage, and errors are captured automatically on every span.

    Per-agent workspace

    Pick an agent and get Overview, Sessions, Traces, and Analysis sub-tabs. Drill from high-level metrics straight into the conversation that caused them.

    Halo agent-loop optimization

    Halo (Hierarchical Agent Loop Optimization) lives inside the Agents dashboard. It reads your traces, finds systemic failure modes across many runs, and writes up concrete fixes with citations back to specific trace IDs.

    Scheduled analysis

    Put Halo on an hourly, daily, weekly, or monthly cadence. Each run lands as a new report so regressions surface before a user does.

    Agent identity that holds up

    Stable agent.id and agent.name attributes group every run cleanly in the dashboard, so sessions stay coherent across deploys and retries.

    Inspect from the CLI

    The same trace data is available from inf trace list, inf trace get, and inf span list. Stay in your terminal when you'd rather grep than click.

    Capabilities

    Every span,
    fully annotated.

    Tracing captures the data you need to debug an agent end-to-end, not just a request log.

    Inputs and outputs

    Every span records the exact data that went in and the exact data that came out, so the input behind any result is one click away.

    Full message history

    User, system, assistant, tool, and tool-result messages preserved in order across the run.

    Tool calls and results

    Tool names, IDs, JSON arguments, and tool outputs captured automatically by the instrumentations.

    Model metadata

    Model name, provider, invocation parameters, finish reason, and prompt/completion/cache token counts.

    Agent structure

    Agent spans, framework spans, tool spans, and graph/node spans. The shape of the run, not just its edges.

    Errors and exceptions

    Failed spans carry exception status and error details so you can find the regression that broke the loop.

    Security

    Enterprise security, built in.

    SOC 2 compliant with encryption at every layer, automatic secret stripping, and full control over data retention.

    Enterprise security built in

    SOC 2 Type II Compliant

    Audited controls across the stack. Meet your security team's requirements without a months-long vendor review.

    Secrets Never Logged

    API keys, tokens, and credentials are detected and excluded from all traces and logs. No configuration required.

    Encryption in Transit & at Rest

    Every request is encrypted end-to-end. Data at rest is encrypted with modern standards so production traffic is never a liability.

    Full Data Retention Controls

    Set retention policies that match your compliance requirements—or turn off data retention entirely. Your data, your rules.

    Trace, analyze, fix, repeat.

    Drop in the SDK, capture your first trace in minutes, and run Halo on real production runs. The whole loop (instrumentation, analysis, and the fixes it surfaces) works on a free account.

    Talk to an EngineerGet Started
    Trace