dropbox-mcp
Exposes the Dropbox API v2 to LLM agents as a fixed set of MCP tools, enabling file management, search, shared links, and account operations with support for Business team-space namespaces.
README
dropbox-mcp
A Model Context Protocol server that gives an LLM agent read/write access to a Dropbox account — personal or Business/Team — over the Dropbox API v2.
The problem
Dropbox's own desktop client solves file syncing. It does not solve giving an AI agent controlled access to an account it doesn't have a local copy of. An agent that can only see synced folders can't search a 2 TB team archive, can't mint a share link, and can't read a file it was never given.
This server closes that gap. It talks to the Dropbox HTTP API directly, so it
works against the whole account without syncing a single byte to disk, and it
exposes that access as a fixed set of MCP tools rather than a raw HTTP client —
the agent gets dropbox_search, not fetch.
Architecture
Four tool modules (files, search, sharing, account) register themselves
against a single McpServer speaking MCP over stdio. Every module funnels
through one HTTP client that owns the two things the Dropbox API makes
annoying: the OAuth token lifecycle and the team-space namespace. Modules can
be disabled individually at startup via DROPBOX_DISABLED_MODULES, so a
deployment that shouldn't be able to create public share links simply doesn't
register that module — the capability is absent, not merely discouraged.
MCP host (Claude, etc.)
| stdio (JSON-RPC)
+-------v--------------------------------------+
| index.ts module registry / stdio transport|
+-------+--------------------------------------+
|
+-------v-------+ +--------+ +---------+ +---------+
| tools/files | | search | | sharing | | account |
+-------+-------+ +---+----+ +----+----+ +----+----+
| | | |
+------+------+-----------+-----------+
|
+--------v-----------------------------------+
| client.ts |
| - refresh-token -> access-token cache |
| - 401 retry with a forced refresh |
| - Dropbox-API-Path-Root resolution |
| - ASCII-safe Dropbox-API-Arg encoding |
+--------+-----------------------------------+
|
RPC api.dropboxapi.com Content content.dropboxapi.com
Two endpoint families, deliberately kept as separate functions: RPC endpoints
are JSON-in/JSON-out, while Content endpoints put their arguments in an HTTP
header and use the body for file bytes. Collapsing them into one generic
request() would have meant a parameter that silently changes where the
arguments go, so they stay apart.
The genuinely hard part
On a Dropbox Business team, the API defaults to the member's home namespace.
Every team folder — which is where the actual shared work lives — sits in the
team's root namespace instead and is simply invisible. list_folder on ""
returns the member's private files and nothing else, with no error and no hint
that most of the account is missing. It looks like a permissions problem and
isn't one.
The fix is to send a Dropbox-API-Path-Root header pointing at the team root
namespace, whose id you get from users/get_current_account. That introduces a
recursion trap: the generic RPC helper attaches the path-root header to every
call, so having it resolve the namespace by calling get_current_account
through itself means the header resolution calls the header resolution
forever. resolveRootNamespaceId() therefore issues a deliberately bare
fetch that bypasses the helper, and the result is memoised so the extra round
trip happens once per process (src/client.ts).
A smaller version of the same class of bug: Content endpoints pass their
arguments in Dropbox-API-Arg, and HTTP headers are ASCII. Any file with an
accented character or an emoji in its name throws inside fetch rather than
returning an API error. apiArg() escapes every non-ASCII code point to
\uXXXX before the header is built.
What I'd do differently
- No tests. This was built and verified by hand against a live account. The token-refresh path, the 401 retry, and the chunked upload boundary conditions are exactly the code that should have been driven by tests with a mocked transport — they're the parts that fail rarely and expensively.
- The path-root cache is per-process and never invalidated. Fine for a stdio server that a host restarts freely; wrong for anything long-lived where a user could be moved between teams.
dropbox_deleteis exposed with no confirmation affordance. Dropbox's own retention makes it recoverable, but a destructive tool should signal that in its schema rather than relying on the host to ask.- Errors are shaped into strings. Returning a structured error code alongside the message would let an agent branch on "rate limited" versus "not found" without parsing prose.
Setup
Requires Node 20+ (developed on 24) and a Dropbox account.
1. Create a Dropbox app
- Go to https://www.dropbox.com/developers/apps and choose Create app.
- Pick Scoped access and Full Dropbox access.
- On Permissions, enable
account_info.read,files.metadata.read,files.metadata.write,files.content.read,files.content.write,sharing.read,sharing.write, then Submit. - On Settings, copy the App key and App secret.
2. Build and authorise
git clone <this-repo>
cd dropbox-mcp
npm install
npm run build
cp .env.example .env # fill in DROPBOX_APP_KEY and DROPBOX_APP_SECRET
npm run auth # prints DROPBOX_REFRESH_TOKEN — paste it into .env
npm run auth prints a consent URL, takes the code you paste back, and
exchanges it for a refresh token. The server trades that for short-lived access
tokens on its own from then on.
3. Register with an MCP host
{
"mcpServers": {
"dropbox": {
"command": "node",
"args": ["/absolute/path/to/dropbox-mcp/dist/index.js"],
"env": {
"DROPBOX_APP_KEY": "your-app-key",
"DROPBOX_APP_SECRET": "your-app-secret",
"DROPBOX_REFRESH_TOKEN": "your-refresh-token"
}
}
}
}
Restart the host and ask it something like "What's my Dropbox space usage?".
To exercise the server without a host:
npm run inspect # @modelcontextprotocol/inspector
Tools
files — dropbox_list_folder, dropbox_get_metadata,
dropbox_create_folder, dropbox_move, dropbox_copy, dropbox_delete,
dropbox_get_temporary_link, dropbox_read_file, dropbox_upload
(auto-chunks large files through upload sessions).
search — dropbox_search across filenames and content, account-wide or
scoped to a folder.
sharing — dropbox_create_shared_link, dropbox_list_shared_links,
dropbox_get_shared_link_metadata, dropbox_revoke_shared_link.
account — dropbox_get_current_account, dropbox_get_space_usage.
Configuration
All configuration is environment variables; see .env.example
for the full annotated list, including the namespace control
(DROPBOX_PATH_ROOT), the read-size cap (DROPBOX_MAX_READ_BYTES), the upload
chunk size, and the team-app impersonation headers.
Licence
MIT — see LICENSE.
Recommended Servers
playwright-mcp
A Model Context Protocol server that enables LLMs to interact with web pages through structured accessibility snapshots without requiring vision models or screenshots.
Audiense Insights MCP Server
Enables interaction with Audiense Insights accounts via the Model Context Protocol, facilitating the extraction and analysis of marketing insights and audience data including demographics, behavior, and influencer engagement.
Magic Component Platform (MCP)
An AI-powered tool that generates modern UI components from natural language descriptions, integrating with popular IDEs to streamline UI development workflow.
VeyraX MCP
Single MCP tool to connect all your favorite tools: Gmail, Calendar and 40 more.
graphlit-mcp-server
The Model Context Protocol (MCP) Server enables integration between MCP clients and the Graphlit service. Ingest anything from Slack to Gmail to podcast feeds, in addition to web crawling, into a Graphlit project - and then retrieve relevant contents from the MCP client.
Kagi MCP Server
An MCP server that integrates Kagi search capabilities with Claude AI, enabling Claude to perform real-time web searches when answering questions that require up-to-date information.
Neon Database
MCP server for interacting with Neon Management API and databases
Exa Search
A Model Context Protocol (MCP) server lets AI assistants like Claude use the Exa AI Search API for web searches. This setup allows AI models to get real-time web information in a safe and controlled way.
Qdrant Server
This repository is an example of how to create a MCP server for Qdrant, a vector search engine.
E2B
Using MCP to run code via e2b.