nduckmink/arkon

git clone https://github.com/nduckmink/arkon.git && cd arkon
cp .env.docker.example .env.docker and set SECRET_KEY, admin credentials, Postgres/MinIO secrets
docker compose --env-file .env.docker up -d --build
Open http://localhost:3119, sign in as admin, and pick embedding/LLM/vision models + paste API keys

Usage

• •Connect Claude Desktop or Claude.ai via a custom connector (URL https://your-arkon-server/mcp) — OAuth 2.1 + PKCE handles sign-in
• •Add 'always search Arkon first using search_wiki' to Claude's Custom Instructions (or Project Instructions) to make Claude use it consistently
• •Browse the wiki (three-panel: page tree, content, backlinks), knowledge graph, full-text + semantic (pgvector) search
• •Ingest documents: reviews a human-reviewable plan before writing pages; merges instead of overwriting existing pages

Key features

• ✓MRP pipeline (Map -> Reduce -> Plan-review -> Refine -> Verify -> Commit): page merge, traceable source claims, image-aware captions, resumable runs
• ✓Department & global scopes with hard enforcement at API/MCP/search layers; fine-grained RBAC (Viewer/Contributor/Editor/Admin) with audit log
• ✓MCP server tools: wiki (search_wiki, read_wiki_page, list_wiki_pages), source drill-down, edit workflow (propose/review/approve/reject drafts), discovery
• ✓AI Skills distribution (versioned .zip SKILL.md packages, department-scoped); catalog-driven LLM/embedding/vision providers
• ✓Self-hosted, no telemetry (outbound only to your AI provider), Fernet-encrypted API keys in PostgreSQL

Best for

Organizations that want a self-hosted, permission-scoped knowledge wiki served to Claude so employees get role/department-filtered internal context.

Caveats

• ⚠License is PolyForm Internal Use 1.0.0 — free for internal business operations but you may not offer Arkon as a service to third parties
• ⚠Built for teams, not individuals (README explicitly suggests Obsidian + Claude Skills for personal setups); runs 7 Docker containers
• ⚠Server requirements: 2-8+ vCPUs, 4-16+ GB RAM (RAM is the primary bottleneck), 40-250+ GB SSD, Ubuntu 22.04+/Debian 12+
• ⚠All AI inference happens externally (Anthropic/Google/OpenAI APIs) — outgoing API keys required; GPU not needed
• ⚠Rich media ingestion (bulk folders, video, Excel) and external data connectors (SharePoint, Google Drive, Notion) are on the roadmap, not yet done