The current state of agentic AI is a mess of fragile RAG pipelines and hallucinated citations. Claix, a new service out of Stockholm, is betting that the missing piece isn't a smarter model, but a robust data and memory layer that treats agents as first-class citizens. The platform recently hit Hacker News with a pitch that focuses less on raw intelligence and more on deterministic, auditable data extraction for autonomous systems.

Agent-to-Agent Protocol Integration

Unlike typical APIs that force developers into proprietary SDKs, Claix is built API-first with native support for the Agent-to-Agent (A2A) protocol. Peers can discover Claix via an Agent Card and invoke over 20 extraction skills using JSON-RPC. This allows orchestration frameworks to treat document processing not as a black box, but as a discoverable service within a multi-agent architecture, maintaining the flexibility of REST for backend integration while speaking the native language of modern agent swarms.

Source Tracing and Hallucination Control

The standout feature here is 'Source Tracing,' which forces every extracted field to return a tuple of { value, source } pointing to a specific page, cell, or fragment. If the system cannot anchor a fact to evidence, it returns requires_human_revision rather than inventing a citation. This 'explainable AI extraction' approach aims to kill the hallucination problem at the data layer, providing an audit trail that backend systems can actually verify, rather than relying on opaque confidence scores.

Technical Architecture and Data Management

The platform uses a clean ID-based architecture where extraction structures are managed via API under a unique schema_id. This replaces 'spaghetti code' and unstable prompts with typed JSON contracts. By associating processed documents with a shared space_id, agents can run cross-document queries to compare or reconcile data across multiple files in a single call, effectively turning static documents into a queryable knowledge base without the maintenance burden of traditional vector databases.

Zero Training and EU Hosting

Privacy is the other major pillar. Claix explicitly states that neither the platform nor its providers use customer documents to train models. Hosted in the EU with encryption at rest and in transit, the service offers granular control over memory retention, allowing users to choose between instant deletion, temporary session memory, or persistent encrypted contexts. This 'zero AI training' stance is a direct counter to the data-grabbing tendencies of major LLM providers, appealing to enterprises wary of leaking IP into foundation models.

Key Takeaways

  • Claix supports A2A protocol discovery via Agent Cards for native agent orchestration.
  • 'Source Tracing' requires evidence-backed extraction, falling back to human review if anchoring fails.
  • Data is hosted in Stockholm with a strict 'zero training' policy on customer documents.
  • The service integrates via REST, MCP for IDEs, and A2A, requiring no mandatory SDKs.

The Bottom Line

This is a necessary infrastructure play. Agents are useless if they can't trust their own memory, and Claix is building the database layer that agents actually need.