three.dev
three.dev scores production AI-agent traffic, groups detected problems into failure modes, and provides evidence for investigation. Engineering and AI teams use it to prioritize fixes and compare agent quality before and after changes.
Who it's for
- AI application teams running agents in production who need to find recurring response failures.
- Engineering teams using Claude Code, Codex, or Cursor who want a coding agent to investigate recorded failure modes in their repository.
- Teams testing agent or prompt changes that want to compare recorded traffic or split live sessions between versions.
How it fits your workflow
Route application requests through the three.dev proxy to an AI provider. As responses return to the app, three.dev sends them to its evaluator for scoring.
Review issues and their evidence, then use grouped and ranked failure modes to decide what to investigate. A coding agent can retrieve the failure details and search the repository for a fix.
Replay recorded traffic or compare traffic slices and versions to see which failure modes changed after a fix.
Pricing
Free: The launch offer provides six months of the Starter plan free for teams who sign up by November 22. The page does not specify the year or what the plan costs after the offer.
Starter
Six months free for teams who sign up by November 22; the page does not state the year or the plan's regular price.
- · Starter plan
- · Six months free under the launch offer
Prices checked on Nov 12, 2026 from the vendor's site. They can change; confirm before you buy.
Key features
- Live traffic scoring
- Scores agent conversations as they pass through the three.dev proxy and sends responses to the app and evaluator at once. The page says scoring adds no latency.
- Evidence-backed issue detection
- Detects issues in agent conversations and attaches evidence, such as the rule violated and the relevant system-prompt line or conversation message.
- Failure-mode grouping and ranking
- Groups individual issues by root cause and displays their frequency and severity to help teams prioritize what to fix.
- Coding-agent repair workflow
- A coding agent can read a failure mode and its recorded evidence, search the repository for the cause, and make a code change.
- Before-and-after comparison
- Compares traffic slices, such as the period before and after a fix or two prompt versions, and shows changes in failure modes.
- Recorded-traffic replay and experiments
- Replays past traffic to test a change without exposing it to users, or splits live sessions between versions and compares quality and business metrics.
Works with
- OpenAI, Anthropic, Gemini, and Azure OpenAI
- Claude Code, Codex, and Cursor
- MCP server
- GitHub skills repository
Limitations to know
- The page offers early access by request and does not state general availability.
- Experiments and specialized language models are labeled early access.
- The launch offer is limited to teams signing up by November 22; the page does not specify the year or the Starter plan's regular price.
Alternatives to consider
-
telemetry.dev
Choose it if you need OpenTelemetry-based traces of model calls, tool steps, and retrievals to investigate latency, token usage, and cost.
Getting started
- Request early access to three.dev.
- Paste the setup prompt into Claude Code, Codex, or Cursor.
- Have the coding agent install the three.dev skills and add three.dev to the repository.
- Route the app's LLM calls through the three.dev proxy and review scored traffic.