KiwiFS: Markdown Filesystem for AI Agents and Teams

What changed

KiwiFS introduces a novel markdown filesystem that transforms plain .md files into a structured, searchable, and versioned knowledge base. This project aims to bridge the gap between the human-readable nature of markdown and the operational needs of AI agents and collaborative teams. Unlike traditional approaches that force a choice between agent-incompatible databases or read-only retrieval layers, KiwiFS makes markdown files the single source of truth, with all other functionalities derived from this core. The system is delivered as a single, zero-configuration binary, simplifying deployment and integration.

Key features include:

Core Functionality

• Searchability: Integrates full-text search (BM25 via SQLite FTS5) and pluggable vector search capabilities (supporting OpenAI, Ollama, ONNX, Cohere, Qdrant, Pinecone, and more).
• Versioning: Implements Git versioning for every write operation, providing an audit trail, blame capabilities, diffs, and point-in-time restore.
• Querying: Offers DQL (Data Query Language), an SQL-like query interface for structured data within frontmatter, supporting operations like TABLE, LIST, COUNT, WHERE, SORT, and GROUP BY.
• Accessibility: Provides multiple access protocols including REST, MCP, NFS, S3, WebDAV, and FUSE.

Agent and Team Features

• Agent Integration: Native integration with Claude, Cursor, and other MCP clients, allowing agents to write to the filesystem directly via MCP, REST, or standard commands like cat.
• Human Interface: Includes an embedded web UI with a block editor, wiki links, backlinks, knowledge graph visualization, and dark mode, all delivered within the single binary.
• Data Import/Export: Supports 19 data importers for various sources (Postgres, MySQL, MongoDB, Notion, Obsidian, CSV, etc.) and offers data export in JSONL/CSV formats, including optional embeddings and link graphs.
• Content Health: Features like stale page detection, broken link checking, orphan detection, contradiction finding, and trust-ranked search help maintain knowledge base integrity.
• Schema Validation: Allows enforcement of structure on writes using JSON Schema.
• Multi-space: Supports multiple isolated workspaces on a single server, each with its own Git repository and search index.
• Webhooks: Enables POST requests to specified URLs (e.g., Slack, CI) on write/delete events, with HMAC signing and retry mechanisms.

Installation is straightforward via package managers (like Homebrew) or a shell script, with quickstart commands provided for initialization and serving the application. Writes can be attributed to specific actors using the X-Actor header in requests, with defined fallbacks for anonymous or unauthenticated writes.

Why it matters for builders

KiwiFS directly tackles the challenge of persistent, structured knowledge for AI agents. Builders can now leverage markdown as a robust backend for agent memory and team knowledge sharing, moving beyond ephemeral chat logs or complex database setups. The native support for MCP and various APIs means agents can interact with this knowledge base seamlessly, enabling more sophisticated agent behaviors and collaborative workflows.

Practical impact

AI builders can integrate KiwiFS into their agent frameworks to provide persistent memory, allowing agents to recall past interactions, learned information, and team knowledge. Teams can use it as a self-hosted, version-controlled alternative to proprietary wiki or knowledge base solutions like Notion or Confluence, with the added benefit of AI agent writability. Developers can explore the extensive import/export options to migrate existing knowledge bases or feed data into the system. The ability to serve data via multiple protocols (REST, NFS, S3) also opens up possibilities for integrating KiwiFS into broader infrastructure.

Caveats and source limits

The provided source is a GitHub repository description and excerpt, detailing features and intended use. Specific performance benchmarks, pricing details (as it appears to be an open-source project), and independent reviews are not available in the provided text. The "fresh release" status indicates it is a recent project, and its long-term stability and scalability under heavy load are yet to be demonstrated through community adoption or further releases. The exact scope of "62 MCP tools" and the depth of integration with each are not fully detailed.