arkon

Arkon is an open-source, self-hosted enterprise-grade knowledge management and MCP (Model Context Protocol) server, designed to establish a governed middle layer between organizational data and AI clients. It runs as a centralized MCP Server, compiling an organization's SOPs, policies, and internal documents into a structured, traceable knowledge wiki, then providing it to Claude and other LLMs through a single permission-scoped endpoint. Sources indicate that AI adoption in most organizations is fragmented—employees copy and paste documents into chatbots, leading to inconsistent context, security risks, and duplicated work; Arkon treats AI as a managed organizational resource, allowing every employee to automatically and securely obtain the correct context according to department and role.

I. Intelligent Knowledge Wiki and the MRP Pipeline Unlike vector databases that only perform chunked indexing, Arkon's MRP pipeline (Map → Reduce → Plan-review → Refine → Verify → Commit) compiles documents into a coherent wiki of interlinked pages. Its features include: plan and review before writing, with each ingestion generating a human-reviewable plan listing which wiki pages will be created or updated, and editors able to regenerate the plan with feedback before pages are written; page merging instead of overwriting, where when a new source touches an existing page, the LLM merges the content to avoid losing old knowledge; traceable claims, with each page recording the source documents from which it was compiled; image-aware, with visual descriptions written into the source text before compilation; and recoverable, with drafts persisted mid-pipeline so that after a crash it can resume without redoing expensive LLM work.

II. Wiki Browser and Knowledge Graph Provides a three-column layout (page tree, content, backlinks and outlinks), full-text and semantic (pgvector) search, interactive knowledge graph visualization (by scope or global), wikilink-style cross-references between pages, per-page version history and rollback, and a workflow of draft proposal → editorial review → approval.

III. Departmental and Global Scopes Supports precise context boundaries, with each scope having its own scoped wiki, document set, and department roster. Department-level isolation for HR, Legal, Engineering, etc., plus a global scope for company-level SOPs, policies, and general knowledge. Scopes are enforced at the API, MCP, and search layers, so members can only see knowledge from their own department or the global scope.

IV. Fine-Grained RBAC Provides role-based access control at the departmental and global levels, with built-in Viewer, Contributor, Editor, and Admin roles to control draft, editing, and settings access; supports granular permissions (such as doc:read:own_dept, wiki:edit:all, org:settings:manage, etc.); and logs an audit trail for every privileged operation (settings changes, plan approvals, role updates).

1. MCP Server for Claude and Other AI Clients Employees connect Claude Desktop or Claude.ai to Arkon via OAuth 2.1 + PKCE; they only need to add the server URL and sign in through the browser, with no manual token copying. The tools exposed by the MCP server include: Wiki tools (search_wiki, read_wiki_page, list_wiki_pages, read_wiki_index); source drill-down (get_source, get_source_outline, get_source_pages, list_sources); editing workflows (propose_wiki_edit, edit_wiki_page, list_pending_drafts, review_draft, approve_draft, reject_draft); discovery (list_knowledge_types, get_knowledge_type_docs). All tools are enforced by token scope (department, knowledge type, source list).

2. AI Skill Distribution Custom agent packages can be uploaded at once and distributed within the organization, with support for versioned skill packages with SKILL.md (.zip), department-scoped visibility, and contribution workflows for end-user updates.

3. Pluggable AI Providers Catalog-driven selection, where administrators choose from a curated list with context window, cost, and capability metadata. LLM support includes Anthropic Claude (Opus / Sonnet / Haiku 4.x), Google Gemini (3.x Pro / Flash / Flash-Lite, 2.5), OpenAI (GPT-5.x, GPT-4.x); embedding support includes Google gemini-embedding-* and OpenAI text-embedding-3-*, switchable with online re-embedding migration (atomically switches the active model after completion, with no zero-result search window); vision support includes Google Gemini Flash and the OpenAI GPT-4o series.

4. Privacy and Security Self-hosted, deployable locally or in a private cloud via Docker; no telemetry, with outbound traffic sent only to the selected AI provider; encryption at rest, with API keys stored encrypted in PostgreSQL using Fernet. Tech stack: backend is FastAPI, PostgreSQL + pgvector, Redis (arq workers), MinIO; frontend is Next.js, Tailwind CSS; AI integration uses MCP via FastMCP; document parsing supports PDF, DOCX, DOC, plain text, URLs, and embedded images.

5. Server Requirements and Quick Start Arkon runs 7 Docker containers (PostgreSQL + pgvector, Redis, MinIO, FastAPI API, 2 ARQ workers, Next.js frontend). RAM is the main bottleneck, as the MRP pipeline worker loads large LLM context windows into memory during wiki compilation; storage scales with the document corpus (pgvector indexes, MinIO file storage, and PostgreSQL WAL logs are the main consumers). All AI inference is performed externally (Anthropic / Google / OpenAI APIs), so no GPU is required. For production, a reverse proxy with SSL is recommended (Nginx / Caddy).

Quick Start (Docker): Prerequisites are Docker and Docker Compose, plus an API key from the selected AI provider (Anthropic, Google, or OpenAI). Clone the repository: git clone https://github.com/nduckmink/arkon.git cd arkon Configure the environment: cp .env.docker.example .env.docker

Edit .env.docker - set SECRET_KEY, admin credentials, and Postgres/MinIO secrets

Start: docker compose --env-file .env.docker up -d --build At http://localhost:3119 access the portal, log in as administrator, then go to Settings to select the embedding / LLM / vision models and paste the API key. For the complete environment variable reference, see the Setup Guide.

Connect Claude: In Claude Desktop or Claude.ai Settings → Connectors, add a custom connector, and fill in the URL with https://your-arkon-server/mcp，点击 After connecting, log in with Arkon credentials in the browser. Arkon uses OAuth 2.1 + PKCE, so no manual token setup is required. To let Claude keep using Arkon, you can add instructions in Claude's Custom Instructions; a stronger constraint is to create a Project in Claude Desktop, attach Arkon as a connector, and paste the same text as Project Instructions.

1. Other Capabilities and Limitations The material also lists the following capabilities (some marked as future or additional features): MRP pipeline, MCP Server, department and global scopes, Wiki drafts and revision workflow, AI skills, catalog-driven model selection, online embedding migration, audit logs, rich media ingestion (including image, video, and Excel table parsing), usage analytics dashboard, external data import (SharePoint, Google Drive, Notion, and other connectors), Arkon CLI, notification system. Arkon uses the PolyForm Internal Use License 1.0.0: free for internal business operations within an enterprise, but Arkon may not be provided to third parties as a service. The material also points out that Arkon is aimed at teams rather than individuals; personal knowledge scenarios are better suited to Obsidian + Claude Skills. The main author is Minh Nguyen (Vietnam, GMT+7). Specific pricing information such as API keys and cost parameters is not yet provided in the material.