OpenCrab
OpenCrab, by Korea's AgentKorea, turns documents into evidence-backed ontology packs queryable via MCP from Claude or Cursor, paid tiers $10-30/month.
Product facts
Commercial Suite- Pricing
- Free tier + subscription
- Data checked
- 2026-10-06
Capabilities
- Graph RAG enhanced retrieval
- Model Context Protocol (MCP) integration
- 9-space meta-ontology extraction
- Pack publishing and marketplace distribution
Snapshot figures reflect the check date and may change over time.
OpenCrab is an ontology-first knowledge organization and graph retrieval platform built by the Korean team AgentKorea, with its site at opencrab.sh running mainly in Korean but switchable to English (checked 2026-10-06). Worth flagging: the commonly indexed domain opencrab.com belongs to an unrelated company’s early “web intelligence aggregation” landing page — a completely different product from the ontology graph engine described here, so don’t confuse the two while researching.
Core Purpose
OpenCrab’s pipeline has four stages: ingest documents, URLs, notes, or code with automatic parsing, OCR, and chunking; extract entities, concepts, evidence, and claims into a sourced ontology graph; package that graph into a reusable “Pack” that teams can share internally or sell on the marketplace; and finally query it with Graph RAG from OpenCrab’s own interface, any MCP client, or a connected AI workspace, getting back answers with a traceable evidence chain. The open-source local engine, LocalCrab (GitHub repo AlexAI-MCP/OpenCrab, 110 stars as of 2026-10-06), implements the MetaOntology OS’s “9-space” grammar (subject, resource, and other dimensions), MCP tools, and Neo4j graph validation, while the hosted opencrab.sh SaaS handles ingestion, the marketplace, and community distribution — that hosted layer is closed source. Compared with Morphik, which embeds pages directly for multimodal vector retrieval without forcing a modeling step, OpenCrab insists on converting material into a structured ontology pack first, trading extra up-front structuring for a traceable, resellable knowledge asset.
Key Scenarios
- Researchers and analysts: turn papers and reports into sourced evidence graphs, replacing “I think I remember” with a verifiable answer chain.
- Domain experts: package specialized knowledge into ontology packs and sell or reuse them through the marketplace.
- AI application builders: expose one graph through a single MCP server to Claude, ChatGPT, Cursor, and Codex alike, as a shared long-term memory layer for agents.
Trying it and its limits
The free tier (checked 2026-10-06) covers browsing public ontology packs, basic graph search, and feeding results to external AI tools; Pro at $10/month adds a private-document personal graph workspace; Expert at $30/month opens pack publishing, management, selling, and MCP reverse-ingestion, with billing through Paddle and plan changes anytime. Two concrete limits: the product and its community run mostly in Korean, so English docs and support are thinner; and ontology modeling requires the ingest-structure-package flow before you get anything queryable, unlike instant vector search, so there’s real setup cost before first use. The private implementation — opencrab.sh’s own ingestion and marketplace backend — stays closed; only the LocalCrab engine is public. Its own navigation bar files the product under a “PET 프로젝트” (personal/side-project) section, a sign this runs more like an independent developer’s product than a well-funded startup — worth weighing against any plan to depend on it long-term.