kiwifs/kiwifs

Markdown filesystem for agents and teams.

What it solves

KiwiFS addresses the problem where raw Markdown files lack the structure, searchability, and versioning required for AI agents and teams to collaborate effectively. It transforms a directory of .md files into a writable, searchable, and versioned knowledge base that remains human-readable and portable, avoiding the lock-in associated with proprietary SaaS databases.

How it works

KiwiFS operates as a single Go binary that manages Markdown files on disk as the single source of truth. It provides a

  • Agent Interface: AI agents can interact with the files via 62 native MCP (Model Context Protocol) tools, REST API, or direct filesystem access (NFS, S3, WebDAV, FUSE).
  • Search Index: It implements full-text search (BM25 via SQLite FTS5) and pluggable vector search (supporting OpenAI, Ollama, ONNX, Cohere, and various vector stores like Qdrant or Pinecone).
  • Versioning: Every write operation is automatically converted into an atomic Git commit, providing a full audit trail and point-in-time restore capabilities.
  • Human Interface: An embedded Web UI featuring a block editor, wiki links, and a knowledge graph visualization for humans to browse and edit.

Who it’s for

  • AI Agent Builders: Developers creating agents that need a persistent, structured memory or workspace where context compounds across sessions.
  • Teams replacing Notion/Confluence: Teams seeking a self-hosted, Markdown-based wiki with no vendor lock-in.
  • Compliance-heavy industries: Organizations requiring an immutable audit trail of all changes via Git.
  • DevOps/Platform teams: Teams maintaining runbooks that agents can update automatically after incidents.

Highlights

  • Native MCP Integration: 62 tools for seamless connection to Claude, Cursor, and other MCP clients.
  • Hybrid Search: Combines full-text and semantic vector search with an option for zero-dependency local ONNX embedders.
  • DQL Queries: SQL-like queries over Markdown frontmatter for structured data retrieval.
  • Zero Config: Distributed as a single binary with an embedded frontend.
  • Multi-space: Support for multiple isolated workspaces, each with its own Git repo and search index.