The Anosys platform
- OpenTelemetry native, plus SDKs for OpenAI Agents, Anthropic, Claude Code and Codex
- LLM judges, datasets, experiments and human review
- Anomaly detection and root cause reports across any dimension you send
- An AI Assistant that queries your data and builds widgets, pipelines and alerts
A trace tells you what one run did. It doesn't tell you that answers got worse after Tuesday's prompt change, that one tool now costs twice as much, or that users stopped finishing checkout after the assistant replied. Those answers need traces, evals, cost and user behavior side by side.
The agent answers, nothing errors, and quality still drops. Judges on sampled traces and user signals such as retries, abandoned sessions and feedback show it before support tickets do.
Retry loops, oversized tool results and model changes raise the bill with no single cause. Cost per run, agent, tool, model, user and developer shows where it went.
The cause can sit in a prompt, a tool, a provider, a data pipeline or the UI. Having all of it joined on the same IDs turns days of stitching queries together into one investigation.
Send what you have
Traces, logs, evals, costs, web events and business events, from any framework.
Join it on shared IDs
Runs, users, sessions, conversations and workflows line up across sources.
See what changed
Eval scores, anomalies and RCA reports point to the slice and the step.
Fix it and check
Alert the owner, ship the change, and compare before and after on the same metrics.
Get data in, shape it, and keep it. Retention and volume limits depend on your plan.
IngestionSend traces, metrics, logs, LLM calls, evals and business events over OTLP, REST, the web tracker, a pixel, the Python and Node.js SDKs, or file drops from Google Cloud Storage and FTP.
Frameworks that emit OpenTelemetry, such as LangChain or CrewAI, work as they are. OpenAI Agents, the Anthropic Agent SDK, Claude Code and Codex have dedicated integrations. Field mapping lets you name custom fields once and use those names everywhere.
Ingestion docs →
Pipelines and process unitsProcess units are reusable steps: a query, a detection algorithm, an extractor, an alert. Pipelines chain them and run on a schedule, with run history and incidents for each one.
Use them to compute a daily metric, flag failed agent traces, summarize refusals, or enrich events with customer and workflow data before they reach a dashboard.
Pipeline docs →Find problems that dashboards miss and narrow them down to a cause.
Anomaly detectionFind cost spikes, latency drift, error bursts and drops in success rate, including the ones that never raise an error.
Detection runs as process units against baselines from your own history, so you can tune it per metric instead of keeping a wall of static thresholds.
Detection docs →
EvaluationsScore production traces and datasets with LLM judges, regex checks or code. Start from templates, version the prompt, criteria and thresholds, and schedule runs.
Compare prompts, models or agents in experiments, track pass rate and judge cost by day, and calibrate judges against human review.
Evaluation docs →
Root cause analysisWhen a metric moves, RCA reports break it down by every dimension you send, such as model, tool, customer segment or client, and rank the slices that changed most.
Open the traces behind a slice to see the exact step, and ask the AI Assistant to summarize what changed and why.
Root cause analysis →Watch the numbers that matter and get told when they move.
Alerts and AutoJudgeCreate alerts with deterministic rules or conditions evaluated by an LLM, and send them to Slack, email or a webhook for your incident tool.
AutoJudge evaluates live traffic against your criteria and can raise an alert when a judge fails, so quality problems page someone the way errors do.
Alert docs →
DashboardsStart from templates for website and product analytics, agent traces, and Claude Code or Codex spend, or build your own with 20+ widget types.
Widgets include funnels, Sankey diagrams, heatmaps, calendar heatmaps, gauges, trace trees and timelines, RCA tables and notes. Each one is built from a query you can open and edit.
Dashboard docs →
AI AssistantAsk questions about your data in plain English. The assistant queries it, explains the result, and can create the widget, pipeline or alert you asked for.
Suggested questions are tied to your datasets, and an assistant widget can sit on any dashboard with the data it should look at already selected.
Assistant docs →From connecting a source to answering a question about it. Names in these screenshots are placeholders.
One endpoint per source
Each source gets its own ingestion endpoint and type: AI agents and OpenTelemetry, website and advertising, or generic events. You can see at a glance which ones are sending data.
- OTLP, REST, web tracker, pixel and SDK ingestion
- Live activity per endpoint
- Field mapping to give custom fields readable names
Follow every agent run, span by span
The timeline shows a multi-agent run as nested spans: agents, turns, model calls, tool functions and handoffs, each with its duration. Open any span for its inputs, outputs, tokens and cost.
- Waterfall of agents, turns, responses and tool calls
- Duration of every span, to the millisecond
- Filter traces and step through records
See the run as a tree
The trace tree shows the structure of a run: where agents branch, where handoffs happen, and which model and tool calls hang off each turn. A retry loop or a dead end is easy to spot in the shape.
- Node graph of agents, responses, functions and handoffs
- Colors by span type
- Works as a dashboard widget
Judges you can version and calibrate
Evaluation templates hold the judge prompt, criteria and thresholds, with versions. Runs score production traces on a schedule, and the overview shows pass rate, judgments, tokens and judge cost by day.
- Simple AI, advanced AI and regex evaluators
- Scheduled eval runs over production traces
- Experiments and human review for calibration
Spend for agents and coding assistants
Token usage and cost per run, model, agent and tool for your own agents, and per developer and project for Claude Code and Codex, from templates you can start with on day one.
- Input, output and cache tokens tracked separately
- Claude Code and Codex templates
- Alerts on daily cost per developer or agent
Product and web analytics next to your AI
The web tracker feeds sessions, page views, engagement and geography into the same workspace as your agent data, so a drop in conversion and a change in agent behavior show up side by side.
- Website, product and agent templates
- Funnels, Sankey diagrams, maps and 20+ other widgets
- Every widget is an editable query
Pipelines and process units
Pipelines run process units on a schedule: compute a metric, flag failed agent traces, summarize refusals, detect anomalies. Each pipeline keeps its run history and incidents.
- Scheduled pipelines with status and history
- Reusable process units for detection, extraction and RCA
- Results to dashboards, Slack, email or webhooks
The AI Assistant
Ask about your data in plain English. The assistant writes and runs the query, explains what it found, and can build the widget, pipeline or alert for you. Suggested questions come from the datasets you already have.
- Grounded in your own datasets
- Builds widgets, pipelines and alerts
- Available as a widget on any dashboard
