Voker Launches Agent Analytics Platform with Intent/Correction/Resolution Primitives
Voker.ai, a YC S24 startup, has launched an analytics platform purpose-built for AI agents. The core product is a lightweight SDK (Python & TypeScript) that wraps LLM calls to OpenAI, Anthropic, and Gemini, automatically collecting conversation data and annotating three primitives: Intents, Corrections, and Resolutions.
What It Does
Voker processes LLM calls by automatically classifying user goals (intents), detecting when users correct the agent (corrections), and measuring when the agent resolves the intent (resolutions). It then uses hierarchical text classification (not LLMs for data engineering) to aggregate these into dynamic categories, giving product teams self-service insights without reading individual traces.
Key Details from the Launch
- SDK Integration: Two lines to install:
pip install vokerand wrapping the LLM provider (e.g.,from voker.ai.provider_openai import OpenAI). - LLM Stack Agnostic: Works with OpenAI, Anthropic, Gemini, Langchain, CrewAI, and Vercel AI SDK.
- Pricing: Free tier — 2,000 events/month (email signup required). Paid plans start at $80/month with a 30-day free trial.
- Data Engineering Philosophy: Voker explicitly avoids using LLMs for core data processing to ensure consistent, reproducible, and accurate statistics. Co-founders note that uploading logs to ChatGPT often yields overfitted or inconsistent insights.
Why It Exists
According to a survey of YC founders, 90%+ said the only way they know agents are failing is via customer complaints. Existing tools fall short: observability (e.g., Langfuse, Langsmith) is good for trace debugging but not accessible to non-engineers; evals test known issues but miss unexpected trends; traditional analytics (PostHog, Mixpanel) aren't built for unstructured conversational data.
Who It's For
Teams running high-volume conversational agents (1k+ chat sessions per month) with complex multi-turn interactions, needing insights that cross-functional teams (PMs, engineers, analysts) can self-serve.
📖 Read the full source: HN AI Agents
👀 See Also
Reverse-engineered Claude Code SDK released in four languages
A developer has reverse-engineered Claude Code and created single-file SDKs in Node.js, Python, Go, and Rust with zero dependencies. The tools provide full agent loop with streaming and tool use while using existing Claude Pro/Max subscriptions.
Benchmark shows AI browser automation tools vary 2.6x in token costs despite identical accuracy
A benchmark of 4 CLI browser automation tools using Claude Sonnet 4.6 on 6 real-world tasks found all achieved 100% accuracy, but openbrowser-ai used 36,010 tokens while others used 77,123-94,130 tokens. Tool call count was the strongest predictor of token cost.
Claude Code's Illusion of Finished Work: Why Reviewing the Agent's Path Matters More Than the Diff
Claude Code can produce a clean diff, passing tests, and a good summary—yet still miss real behavior, security concerns, or architecture constraints. The author argues that reviewing the chain of actions (plans, files read, commands run, test output) is now essential, not just the final diff.
Automated .xcstrings Localization with Claude Code
A new Claude Code skill automates localization of Xcode .xcstrings files with five pipeline stages: domain scanning, comment generation, translation with CLDR plurals, grammar checking, and plural fix conversion.