arkon

Arkon is a self-hosted knowledge layer for organisations that sits between internal documents and AI clients. Rather than chunking files into a vector index, its MRP pipeline (Map, Reduce, Plan-review, Refine, Verify, Commit) compiles documents into a wiki of interlinked pages. Each ingestion produces a plan listing which pages will be created or updated that a human can review first; new material is merged into existing pages instead of overwriting them; every page records its source documents; and drafts persist so a crashed run resumes without repeating the LLM work.

People then use the wiki with a browser view (page tree, backlinks, a knowledge graph, full-text and pgvector semantic search, version history with rollback) or from Claude Desktop and Claude.ai through the MCP server, which signs in with OAuth 2.1 and PKCE. Its tools cover wiki search and reading, source drill-down, and an edit workflow with propose, review, approve and reject. Access is controlled by department and project workspaces, roles (Viewer, Contributor, Editor, Admin and custom ones) and an audit log, with scope enforced at the API, MCP and search layers. It can also distribute versioned skill packages.

It runs as 7 Docker containers (PostgreSQL with pgvector, Redis, MinIO, the FastAPI API, two workers and a Next.js frontend). The README suggests 2 vCPU and 4 GB RAM for up to 20 people. LLM, embedding and vision calls go to Anthropic, Google or OpenAI with keys you supply, so no GPU is needed, and it says there is no telemetry. Licensing needs attention: it uses the PolyForm Internal Use License 1.0.0, which is free for internal business operations but does not let you offer Arkon as a service to others, so it is source-available rather than open source. Last push was 2026-06-03. This review is based on the README and repository data; we have not run it ourselves.