Anchor Open Protocol Builds Local-First Semantic Models for AI Agents

Summary

Anchor is an open protocol, CLI, and MCP server for building a shared semantic layer above an organization’s existing data sources. It maps tables and documents to entities, properties, relations, and business rules, producing a versioned, committable model.yaml rather than moving or replacing source data. The pipeline separates evidence, inferred meaning, and human-approved truth: local profiling computes statistics, an inference endpoint receives only compressed statistics and patterns, and reviewers arbitrate uncertain proposals. Raw rows, cell values, and document text remain on the organization’s machine; the only inference network call carries column names, types, distributions, and patterns. Six of eight stages use no model call, including ingestion, profiling, review, and serving. Confidence thresholds determine whether items are confirmed, sent for review, or recorded as doubts, while provenance links each model element to source evidence. Content hashes support incremental reruns, cache results, and show ontology changes between runs. The local deployment serves five deterministic MCP operations from model.yaml without an account or runtime network connection; optional telemetry sends only operation names. A hosted worker-service variant accepts multi-tenant uploads, deletes raw files after each run, retains derived artifacts, and records deletions in an append-only audit log. The repository requires Node.js 22 or later, pnpm, and a Vercel AI Gateway key for inference, and publishes a TypeScript SDK for its hosted API.