disk
Disk is a personal cloud storage on Cloudflare with an MCP server enabling agents to perform file operations, sharing, and key management via natural language.
README
<div align="center"> <img src="web/public/assets/logo.png" alt="Disk" width="96" height="96" />
Disk
Personal cloud storage on Cloudflare — for you and your agents.
Any file, any size · scoped API keys · full audit trail · agent-ready (REST + OpenAPI + MCP)
disk.zaynjarvis.com · built on Cloudflare Workers · R2 · D1
</div>
<div align="center"> <img src="docs/screenshots/files.png" alt="Disk file browser" width="820" /> </div>
What it is
Disk is one Cloudflare Worker in front of R2 (file bytes), D1 (keys, shares, audit), and a React SPA. It's a Dropbox-lite you own outright — and every capability is equally available over a clean API so your agents can use it as scratch space, artifact storage, or a shared drive.
Features
- 🔑 API-key auth with roles — a
ROOT_KEYenv secret, revocable root-equivalent admin keys, and scoped user keys (path prefix, read-only/read-write, per-file upload cap, expiry). - 🗂️ Virtual root per key — a scoped key sees its prefix as the root and
never perceives or escapes it. Hand an agent a key scoped to
agents/zouk/and, to that agent, its world simply starts at/. - 📤 Any file, any size — drag-and-drop and folder uploads, multipart for multi-GB files, Range downloads, previews (image / video / audio / PDF / markdown / code).
- 👀 Four ways to browse — List, Grid, Tree, and a Finder-style Column view that previews the selected file in a pane on the right, making full use of wide screens.
- 🔗 Shares — revocable, expiring public links (
/s/<token>), download-counted. - 📜 Full audit — every request logged (actor, action, path, bytes, status, IP, UA); filterable UI + CSV export for admins.
- 🤖 Agent-ready — REST at
/api/v1, OpenAPI 3.1 at/openapi.json, an agent guide at/llms.txt, and an MCP server atPOST /mcp— including admin tools so an agent can manage keys itself.
Screenshots
| Sign in | Preview anything |
|---|---|
| <img src="docs/screenshots/login.png" width="400" /> | <img src="docs/screenshots/preview.png" width="400" /> |
| API keys — roles, scopes, upload caps | Storage usage |
|---|---|
| <img src="docs/screenshots/keys.png" width="400" /> | <img src="docs/screenshots/usage.png" width="400" /> |
Column view — Finder-style Miller columns with a live preview pane on the right; and a Tree view for a bird's-eye look. Different views for different audiences.
| Column view + preview pane | Tree view |
|---|---|
| <img src="docs/screenshots/column.png" width="400" /> | <img src="docs/screenshots/tree.png" width="400" /> |
Virtual root — the same disk, seen through a key scoped to agents/zouk/.
No prefix, no other folders, no admin views: its scope is the root.
<div align="center"> <img src="docs/screenshots/scoped.png" alt="Scoped key virtual root" width="700" /> </div>
How to use it
As a person
- Open disk.zaynjarvis.com.
- Paste an API key (your
ROOT_KEY, or any key you minted) and sign in. - Drag files in, preview them, share links, and — as root/admin — mint keys and read the audit log.
As an agent
Mint a key for the agent in the Keys tab (or via the API), then point it at Disk. Give it a scoped key so it can only touch its own corner:
# (admin) mint a scoped read-write key with a 10 MB per-file cap
curl -s -X POST https://disk.zaynjarvis.com/api/v1/keys \
-H "Authorization: Bearer $ROOT_KEY" -H 'Content-Type: application/json' \
-d '{"name":"zouk-agent","prefix":"agents/zouk/","permission":"rw","maxUploadBytes":10485760}'
# → returns {"key":"dk_…"} (shown once)
The agent then speaks scope-relative paths — it never sees agents/zouk/:
export DISK_KEY=dk_...
# upload (stored at agents/zouk/notes/todo.md, but the agent just says "notes/todo.md")
curl -s -X PUT "https://disk.zaynjarvis.com/api/v1/files/blob?path=notes/todo.md" \
-H "Authorization: Bearer $DISK_KEY" --data-binary @todo.md
# list its root
curl -s "https://disk.zaynjarvis.com/api/v1/files" -H "Authorization: Bearer $DISK_KEY"
# share a file
curl -s -X POST https://disk.zaynjarvis.com/api/v1/shares \
-H "Authorization: Bearer $DISK_KEY" -H 'Content-Type: application/json' \
-d '{"path":"notes/todo.md","ttlSeconds":3600}'
Full curl reference for every operation: /llms.txt.
Over MCP
Connect any MCP client (Claude Code, Claude Desktop, Cursor, …). The agent gets
native tools — disk_list, disk_read, disk_write, disk_delete,
disk_move, disk_share, and (for admin keys) disk_key_create /
disk_key_list / disk_key_revoke:
{ "mcpServers": { "disk": {
"type": "http",
"url": "https://disk.zaynjarvis.com/mcp",
"headers": { "Authorization": "Bearer dk_..." }
} } }
Hand an agent an admin key and it can administer the whole disk — including minting and revoking other keys — entirely through MCP.
Stack
Hono on Cloudflare Workers · R2 · D1 · React 19 + Vite + Tailwind v4 · vitest
(@cloudflare/vitest-pool-workers). One wrangler deploy ships the API, the
MCP server, and the SPA together.
Contributing & deploying
- DEV.md — local setup, the contract-first rule, security-core invariants, recipes, and gotchas.
- DEPLOY.md — architecture, provisioned resources, redeploy, migrations, disaster-recovery bootstrap.
- API.md — the binding API contract.
Screenshots above are generated with node scripts/screenshots.mjs against a
local dev server.
推荐服务器
Baidu Map
百度地图核心API现已全面兼容MCP协议,是国内首家兼容MCP协议的地图服务商。
Playwright MCP Server
一个模型上下文协议服务器,它使大型语言模型能够通过结构化的可访问性快照与网页进行交互,而无需视觉模型或屏幕截图。
Magic Component Platform (MCP)
一个由人工智能驱动的工具,可以从自然语言描述生成现代化的用户界面组件,并与流行的集成开发环境(IDE)集成,从而简化用户界面开发流程。
Audiense Insights MCP Server
通过模型上下文协议启用与 Audiense Insights 账户的交互,从而促进营销洞察和受众数据的提取和分析,包括人口统计信息、行为和影响者互动。
VeyraX
一个单一的 MCP 工具,连接你所有喜爱的工具:Gmail、日历以及其他 40 多个工具。
graphlit-mcp-server
模型上下文协议 (MCP) 服务器实现了 MCP 客户端与 Graphlit 服务之间的集成。 除了网络爬取之外,还可以将任何内容(从 Slack 到 Gmail 再到播客订阅源)导入到 Graphlit 项目中,然后从 MCP 客户端检索相关内容。
Kagi MCP Server
一个 MCP 服务器,集成了 Kagi 搜索功能和 Claude AI,使 Claude 能够在回答需要最新信息的问题时执行实时网络搜索。
e2b-mcp-server
使用 MCP 通过 e2b 运行代码。
Neon MCP Server
用于与 Neon 管理 API 和数据库交互的 MCP 服务器
Exa MCP Server
模型上下文协议(MCP)服务器允许像 Claude 这样的 AI 助手使用 Exa AI 搜索 API 进行网络搜索。这种设置允许 AI 模型以安全和受控的方式获取实时的网络信息。