next-wiki

A self-hosted, AI-native wiki for turning conversations and sources into durable knowledge.

Capture what you learn, keep the evidence, ask AI to retrieve and organize it, and publish a readable wiki when it is ready. next-wiki is built for people who want an AI memory they can run, inspect, export, and connect to any compatible client.

GitHub strips <iframe>/<script> tags from rendered READMEs, so this page can only link out rather than embed the demo inline. The project home page embeds a real, live instance you can click through — not a recorded walkthrough — running in read-only demo mode (see NEXT_WIKI_DEMO_READONLY).

Project status: next-wiki is actively developed and currently in an early open-source release. Interfaces and configuration may evolve.

AI administration	Public REST API reference

From capture to durable knowledge

Solid blue represents capabilities available today. Amber dashed cards represent planned import connectors. The editable source and regeneration instructions live in the diagram source.

Why next-wiki?

Most wikis are good at storing pages, while most AI assistants are good at answering questions. next-wiki joins the two into a governed knowledge loop:

conversation / source / command output
              │
              ▼
       raw, append-only evidence
              │
              ▼
    AI retrieval, synthesis, and drafts
              │
              ▼
       human review and publication
              │
              ▼
       durable, searchable wiki memory

The goal is not to hide an AI agent behind a chat box. The goal is to make knowledge useful to both people and agents while keeping ownership, evidence, permissions, revisions, and publication boundaries visible.

Highlights

A complete Markdown wiki

• Edit Markdown with a split source/preview editor and a familiar page tree.
• Render GitHub-Flavored Markdown, syntax-highlighted code, math with KaTeX, Mermaid diagrams, images, frontmatter, tags, and metadata.
• Save drafts, publish explicit revisions, compare any two revisions, and soft-delete pages without losing their history.
• Navigate backlinks, outbound links, page neighborhoods, tags, and related content from the reader.
• Export page Markdown directly, or use the versioned ZIP transfer flow for published pages and referenced local images.

AI that can read, reason, and help curate

• Persistent Wiki AI chat is grounded in the wiki through hybrid keyword and semantic retrieval, with visible sources and retrievable citations.
• Chat sessions are retained, shareable through their URL, and restorable with their conversation context.
• Configure chat, embedding, and image capabilities independently. The admin surface manages providers, models, capability detection, assignments, prompts, runtime parameters, usage, and vector index rebuilds.
• Use AI for page drafting, text improvement, image generation, and queued multilingual translation workflows with immutable provenance.
• When a selected model cannot use tools, Wiki AI degrades to ordinary question answering instead of silently pretending it performed a mutation.

A governed Wiki AI tool runtime

Wiki AI can use a built-in, MCP-compatible next-wiki tool provider to work with the same permission-checked services used by the web UI and REST API. The runtime provides:

• read tools for search, page retrieval, page listing, tags, backlinks, and page neighborhoods;
• draft and organization tools for creating/saving page drafts, changing metadata and properties, managing tags, and proposing batch operations;
• a bounded, provider-agnostic tool loop with live tool-call status in chat;
• server-enforced risk, permission, retention, and review policies;
• Admin proposal review with approve, reject, apply, conflict detection, and per-item results;
• Tool Evidence Raw entries when tool output becomes durable source material;
• audit events and normal page revision/publication boundaries for durable changes.

Only the built-in wiki provider is enabled inside Wiki AI today. External MCP providers are intentionally not auto-discovered or activated in this phase. The separately packaged MCP server is available when an external AI client should connect to next-wiki directly.

A self-growing memory model

Choose the authoring model that fits the instance:

Mode	Best for	Content model
Copilot	A conventional collaborative wiki	Human and AI work in the default wiki space; drafts and publication remain the primary workflow.
LLM Wiki	Evidence-first personal or team memory	raw stores append-only source material, generated stores AI-produced concepts with provenance, and default remains the curated public wiki.

This is the beginning of the README. Read the rest at the original. Read the full README ↗