What it does Arkon ingests your org’s PDFs, policies, and docs through a multi-stage “MRP pipeline” (Map → Reduce → Plan-review → Refine → Verify → Commit) that builds a structured, interlinked wiki rather than a loose bag of vector chunks. It then exposes that wiki to Claude and other LLMs via an MCP server with OAuth 2.1 + PKCE login, scoped by department and role.

The interesting bit The plan-review step is the unusual part: before any wiki page gets written, the system generates a human-reviewable plan showing exactly which pages will be created or updated. Editors can reject and regenerate. When sources overlap with existing pages, content gets LLM-merged rather than overwritten, so knowledge accumulates instead of colliding. The whole pipeline is resumable if a worker crashes mid-compilation.

Key highlights

• Department and global scopes with hard isolation—HR sees HR, Engineering sees Engineering, everyone sees company-wide SOPs.
• RBAC with audit logging built in (Viewer → Contributor → Editor → Admin), plus per-token scope enforcement at the MCP tool layer.
• Online embedding migration—switch embedding models atomically without a zero-result search window.
• 7 Docker containers (FastAPI, Next.js, PostgreSQL+pgvector, Redis, MinIO, 2 ARQ workers); no GPU needed since inference is external.
• PolyForm Internal Use License—free for internal ops, no SaaS resale.

Caveats

• Not for individuals. The README explicitly points personal users to Obsidian + Claude Skills.
• RAM is the bottleneck. The MRP workers load large LLM context windows; 4 GB for starter, 8 GB for teams, 16+ GB for enterprise.
• Several roadmap items are unchecked: rich media ingestion, external data connectors (SharePoint, Google Drive, Notion), CLI setup, and notifications are all pending.

Verdict Worth evaluating if you’re a 20+ person org already paying for Claude/GPT/Gemini and tired of employees copy-pasting inconsistent context into chatbots. Skip it if you’re a solo developer or want turnkey cloud SaaS—this is deliberately self-hosted, Docker-heavy, and team-oriented.