TroveFiles for Claude Desktop, Cursor, and any MCP client.
Persistent project workspace for your AI coding agent. One CLI command wires TroveFiles into every MCP-compatible client on the machine. The server runs locally over stdio — no remote MCP gateway between you and the model.
One command,
every client.
No JSON editing. The CLI auto-detects Claude Desktop, Cursor, and (with --client claude-code) Claude Code, then merges a server entry into each config atomically. Existing MCP servers in the same file are preserved.
# One-time install, on any machine you use Claude Desktop / Cursor / Claude Code
pip install 'trove-sdk[cli,mcp]'
# Log in once with the key from trovefiles.dev/dashboard
trove login --api-key trove-sk-... --namespace my-project
# Detect every MCP client on this machine and wire it up
trove mcp install
# ✓ Claude Desktop: wrote 'trove' → ~/.../claude_desktop_config.json
# ✓ Cursor: wrote 'trove' → ~/.cursor/mcp.json
# restart the client(s) above to load the new server.Three tools,
every Unix command.
Models degrade as the tool list grows. trove_exec is the universal hammer — anything you'd type at a shell prompt — and read/write skip the quoting tax for the common case.
trove_exec(command, stdin?)
Run any shell command in your workspace. Real Unix tools preinstalled:
jq, awk, sed, grep, pdftotext, ffmpeg, imagemagick, python3, curl, …
trove_read(path)
Read a UTF-8 text file from the workspace (1 MB cap).
trove_write(path, content)
Write a UTF-8 text file. Parent dirs created as needed.TroveFiles MCP
or local filesystem?
TroveFiles MCP wins when…
- • You want the same workspace from multiple machines.
- • You need real Unix tooling (jq, pdftotext, ffmpeg) without installing it everywhere.
- • Each project gets its own isolated namespace by default.
- • You want webhooks on every file write — pipe agent activity into your own systems.
- • You'll later embed the same workspace in a SaaS product through the SDK.
Native filesystem access wins when…
- • You only ever use one machine and want zero cloud dependency.
- • Your files must stay on local disk for compliance.
- • You don't need cross-session, cross-machine persistence.
The two are not exclusive — many users keep Claude Desktop's local filesystem access on for personal files and use TroveFiles for project workspaces that follow them between machines.
MCP server,
answered.
Does the MCP server run locally?
Yes. The MCP server is a stdio subprocess your AI client (Claude Desktop, Cursor, etc.) launches on your own machine. There is no remote MCP gateway — your prompts never travel through anyone's MCP relay. The server then makes API calls to TroveFiles's storage backend on your behalf, the same way any local app talks to a cloud database.
Which clients are supported?
Anything that speaks the Model Context Protocol. trove mcp install auto-detects Claude Desktop and Cursor today, and writes a project-scoped .mcp.json for Claude Code when you pass --client claude-code. Other MCP clients (Continue, Zed, custom hosts) work too — copy the JSON entry the install command writes and paste it into the client's config.
How is this different from giving Claude Desktop filesystem access?
Claude Desktop's built-in filesystem access targets files on your local disk. TroveFiles gives you a remote workspace that survives across machines, ships with Unix tooling preinstalled (jq, pdftotext, ffmpeg…), and isolates each project under a separate namespace. The data is in the cloud, available wherever you log in — your laptop, a different machine, your phone — and it can be wired into your own backend through the same SDK.
Can I use a different namespace per project?
Yes. Run trove mcp install --namespace project-foo from inside each project, or pass --client claude-code to drop a .mcp.json into the project root with that namespace baked in. Switch projects, switch namespaces — the agent only sees the workspace for the project it is in.
How do I remove the server?
trove mcp uninstall removes the TroveFiles server entry from every detected client. It only touches the trove key in mcpServers — any other servers you have configured are left alone.
Why only three tools?
Models lose accuracy as the tool count grows. trove_exec is a universal hammer (any Unix command in the workspace), and trove_read / trove_write skip the shell-quoting tax for the common case. Listing files, deleting files, snapshots — all reachable through trove_exec without bloating the tool surface.
Plug TroveFiles into
Claude in 30 seconds.
pip install 'trove-sdk[cli,mcp]' · trove login · trove mcp install. Restart your client and your agent has a persistent workspace.