three.dev
Scores live AI agent traffic, groups failures into failure modes, and hands fixes to your coding agent.
At a Glance
Permanent free plan with room for about 50,000 requests a month at 5k tokens each.
Engagement
Available On
Alternatives
Listed Nov 2026
About three.dev
three.dev is a product from round3 that detects failures in AI agents running in production. It scores live conversations, groups issues into failure modes, and lets a coding agent such as Claude Code, Codex, or Cursor apply fixes, then compares traffic before and after.
What It Is
three.dev is an agent quality and observability tool. Your app's LLM calls are routed through a three.dev proxy to providers such as OpenAI, Anthropic, Gemini, or Azure OpenAI. Each response goes to your app and to an evaluator at once, so the site says scoring adds no latency. Issues are detected automatically with evidence attached, such as the rule violated and the system prompt line that supports it.
From failures to fixes
Issues are grouped by root cause into failure modes, ranked by severity and frequency, with counts of how often each occurs. A coding agent can read a failure mode through the MCP server (list_failure_modes, get_failure_mode, get_request_conversation) and propose a code change. Teams can merge or reshape failure modes, ask questions in a Chat interface, and compare any two slices of traffic, such as before and after a fix or two prompt versions, with new failure modes shown alongside.
Setup path
Setup is one prompt pasted into a coding agent, which installs the three.dev skills from the round3ai GitHub repository, finds LLM calls, routes them through the proxy, and names the use case.
Early access features
The site lists Experiments (replaying v4 on recorded traffic or splitting live sessions) and Specialized Language Models trained on your own traffic as early access.
Community Discussions
Be the first to start a conversation about three.dev
Share your experience with three.dev, ask questions, or help others learn from your insights.
Pricing
Free
Permanent free plan with room for about 50,000 requests a month at 5k tokens each.
- Issues with evidence
- Failure modes
- Your coding agent, over MCP
- Chat
- 1 GB of data a month (Upgrade to send more)
Starter
Room for about 250,000 requests a month at 5k tokens each. Launch offer: free for 6 months for teams who sign up by November 22.
- Issues with evidence
- Failure modes
- Your coding agent, over MCP
- Chat
- 5 GB of data a month, then $3 per GB
- 10k requests scored a month, then $2.50 per 1k
- 90 days of history
- Unlimited seats
- Sampling you control
- Launch offer: Starter free for 6 months (sign up between November 9 and 22 and send first request within a week)
Pro
Room for about 1,000,000 requests a month at 5k tokens each.
- Issues with evidence
- Failure modes
- Your coding agent, over MCP
- Chat
- 20 GB of data a month, then $2 per GB
- 25k requests scored a month, then $1.50 per 1k
- 180 days of history
- Unlimited seats
- Sampling you control
Enterprise
Built for high volume with custom limits, sized to your traffic.
- Issues with evidence
- Failure modes
- Your coding agent, over MCP
- Chat
- Custom data a month
- Custom scoring a month
- Custom history
- Unlimited seats
Capabilities
Key Features
- Live scoring of production agent conversations
- Evidence-backed issue detection
- Failure mode grouping with severity ranking
- Fix with your coding agent
- MCP server integration
- Before and after comparison
- Merge and manage failure modes
- Chat over traffic
- Experiments (early access)
- Specialized language models (early access)
