2026年おすすめ AI Virtual File System オープンソースプロジェクト選定ガイドと導入チェックリスト

一言で： AIエージェントがGitHub Issue、S3ログ、Slackスレッド、ローカルコードを一度に扱うとき、ボトルネックはモデルではなく断片化されたコンテキストIFです。AI VFSはすべてをパスツリーに畳み込み、read・grep・listでクロスソース操作を可能にします。

なぜエージェントに仮想ファイルシステムが必要か

In 2025–2026, AI agent architectures evolved from single-turn Q&A to long-running tasks with multi-tool orchestration. A typical coding agent in one task might:

• Read local src/ TypeScript files
• Pull GitHub PR diffs and CI logs
• Query Postgres user tables for data fixes
• Reply with progress in Slack threads

If every source is a separate MCP server or REST API, the agent's context window fills with tool schemas and auth headers. The core VFS value: fold everything into paths.

Unix philosophy revived for agents: Plan 9 said 'everything is a file'; 2026 AI VFS says 'everything is context'—database tables are directories, API responses are files, Git commits are snapshots.

2026 Landscape: 6 Open-Source Projects Worth Watching

Ranked by maturity, community activity, and use case—all actively maintained as of August 2026.

1. Mirage — The Swiss Army Knife of Unified Data Planes

Repo: strukto-ai/mirage · License: Apache 2.0 · Lang: Python / TypeScript

Mirage is among the loudest AI VFS projects (3k+ GitHub stars). It mounts S3, Google Drive, Gmail, Slack, Redis, Postgres, GitHub, Notion, and 50+ backends under one virtual tree—agents pipe across sources with bash cat, grep, and pipes.

• Highlights: Zero new vocabulary—if you know bash, you're ready; FUSE, MCP, LangChain / Vercel AI SDK
• Unique: Reading PDFs returns parsed pages, not raw bytes; per-resource read semantics
• Best for: Multi-SaaS agent orchestration, RAG pipelines, OpenHands / Claude Code plugins
• Caveat: More backends = more mount config; split root namespaces per tenant in production

2. AgentFS (Turso) — SQLite-Powered Sandbox Filesystem

Repo: tursodatabase/agentfs（Factory-AI fork 活跃维护）· License: MIT · Lang: Rust

AgentFS takes another path: not 50 SaaS connectors, but an auditable, rollback-capable agent workspace. SQLite stores all file content and metadata with copy-on-write isolation and FUSE (Linux) / NFS (macOS) mounts.

• Highlights: Version every agent operation; SQLite transactions for atomicity; multi-agent parallel writes
• Best for: CI agent sandboxes, coding agents needing checkpoint rollback, compliance audit
• Caveat: Linux is first-tier; macOS uses NFS + Seatbelt—run manual validation scripts

3. AFS (AIGNE) — The 'Everything Is Context' Abstraction

Repo: AIGNE-io/afs · License: Apache 2.0 · Lang: TypeScript

AFS defines 8 unified operations: read、write、list、search、stat、exec、explain、delete. Any data source exposes paths via Provider plugins—Git branches as directories, SQLite tables as directories, rows as files.

• Highlights: Clear provider ecosystem (@aigne/afs-git、@aigne/afs-sqlite、@aigne/afs-mcp); AFS-UI lets agents render web pages
• Best for: TypeScript teams needing Git + DB + local files in one internal agent
• Caveat: Beta (v1.11.x), APIs may change; best for greenfield, not large migrations

4. agent-fs — Agent Long-Term Memory with Semantic Search

Repo: desplega-ai/agent-fs · License: MIT · Lang: TypeScript

agent-fs is a persistent agent filesystem + vector index. Beyond CRUD: semantic search (OpenAI / Google / local llama.cpp), DuckDB SQL over documents, S3 sync, and MCP.

• Highlights: Single-binary CLI + HTTP server; Identity files for cross-session agent identity
• Best for: Multi-agent swarm shared workspace, long-term 'what did we write before' memory
• Caveat: New community (Q1 2026); production stories still accumulating; budget embeddings separately

5. VFS (ClayGendron) — Four Verbs: glob / grep / glean / graph

Repo: ClayGendron/vfs · License: Apache 2.0 · Lang: Python

This project abstracts knowledge-base retrieval into four verbs: glob (pattern match), grep (full-text), glean (vector), graph (relations). v2 core router has 1,600+ tests; SQL backends migrating.

• Highlights: Direct DB BM25 + vector + graph; vfs-py on PyPI
• Best for: Enterprise knowledge RAG, deep retrieval agents on SQL backends (Postgres/MSSQL)
• Caveat: Alpha; v2 API incompatible with older PyPI; pin versions and watch CHANGELOG

6. OpenHands Workspace — Runtime Filesystem for Coding Agents

Repo: All-Hands-AI/OpenHands · License: MIT · Lang: Python

OpenHands doesn't call itself a VFS, but its Docker sandbox + workspace mount is the coding agent filesystem layer: isolated read/write, bash, web browsing. Mirage ships an OpenHands adapter.

• Highlights: Mature coding agent runtime; deep GitHub/GitLab integration
• Best for: Fully automated PR fixes, issue-to-code pipelines
• Caveat: Sandbox is Docker-level, not VFS-level; cross-SaaS needs Mirage or MCP

比較表

選定決定木：どれを選ぶ？

1. Primary need: connect 10+ SaaS/APIs? → Mirage. One mount command for S3 + Slack + GitHub.
2. Need agent sandbox isolation + rollback? → AgentFS. SQLite COW is the cleanest 2026 option.
3. Full-stack TypeScript, unified Git + DB access? → AFS. Clear provider model for internal tooling.
4. 多 Agent 共享记忆 + Semantic search？ → agent-fs. Vector index and Identity are differentiators.
5. Enterprise knowledge RAG, deep SQL retrieval? → ClayGendron VFS. Four verbs cover the full retrieval spectrum.
6. Fully automated coding agent (PR fixes)? → OpenHands + Mirage for the data plane.

MCPとVFSの組み合わせ

2026 best practice isn't 'MCP or VFS'—it's layering:

Agent client (Claude Code / Cursor / Codex)
  ├── MCP Layer：Expose VFS as mcp://fs/* tools
  └── VFS Layer：Mirage / AFS / agent-fs Unified data plane
        ├── /local/src          → Local代码
        ├── /s3/logs/           → object storage
        ├── /github/acme/repo   → GitHub API
        └── /pg/users           → Postgres tables

Benefit: the client maintains one MCP connection; mounts, auth, and caching live in the VFS layer. See our MCP解説 article on data source configuration—VFS is essentially MCP's heavy data-plane implementation.

Cloud MacでのAgent VFS運用

大多数 VFS Project的开发和验证环境是 macOS 或 Linux。几个工程要点：

• FUSE limits: macOS FUSE needs macFUSE; AgentFS on macOS is more stable via NFS. Linux cloud VMs run FUSE without friction.
• Persistent mounts: Agent tasks run for hours; laptop lids kill connections. Cloud Mac tmux + fixed IP keeps VFS mounts 24/7.
• Secrets: Don't put 50-backend OAuth tokens in prompts—use VFS-layer .env or Vault sidecar; agents only see paths.
• 与Local推理共存： Mirage for data + on-box Ollama/MLX inference is the best-value private agent stack in 2026.

If you're building an AIコーディングワークフロー, add VFS mount to Cursor Rules or Agent Skills init—auto-mount data sources each session.

7ステップ導入チェックリスト

1. 列出 Agent 需要访问的数据源（Local / S3 / Git / DB / SaaS）
2. 按决策树选定 VFS Project，Local pip install 或 cargo install 验证
3. Configure minimal mount set (2–3 backends first, not 50 at once)
4. Connect via MCP or FUSE to Claude Code / Cursor; run an end-to-end task
5. Load test: 1000 read + grep for latency and token cost
6. 加鉴权隔离：按Project/租户拆分 namespace
7. Move to a persistent Cloud Mac node; hook CI webhooks to trigger agent tasks

Bottom line: AI VFS in 2026 has moved from PoC to production-ready selection. Multi-source orchestration: Mirage. Sandbox isolation: AgentFS. Long-term memory: agent-fs. Deep retrieval: VFS—no single winner, only the best fit for your agent architecture. Unify the data plane first; then argue about model intelligence.