MCP server

Give Claude, ChatGPT or any MCP client your drive through the connector, or run the server on your machine with an API token.

Endpoint
mcp.freakingfast.io/mcp
Transport
Streamable HTTP
Sign-in
OAuth 2.1

Connect a client#

  1. Add the connector

    Connector URL
    https://mcp.freakingfast.io/mcp

    Claude has it in its connector directory: connect it there, and it follows you into the Claude apps. In ChatGPT, add the URL as a connector in settings. Any other client that takes a remote MCP server takes the same URL.

    In Claude Code, run this, then /mcp to sign in.

    Claude Code
    claude mcp add --transport http freakingfast https://mcp.freakingfast.io/mcp
  2. Sign in and choose what it can do

    Your browser opens the drive, where you pick how much the app reaches, where, and for how long. Nothing is let in until you allow it.

  3. Ask

    The tools a connection can't use aren't offered to it at all, so the model never reaches for something it would be refused.

What you allow#

Each level adds to the one before it. An app that asks for less can't be given more.

LevelScopeWhat it can do
Readfiles:readWhat it can doFind, open and link to your files, and see who opened them.
Organizefiles:organizeWhat it can doAlso tag, move, add and arrange files. Nothing it does shows to anyone else, and nothing gets deleted.
Sharefiles:shareWhat it can doAlso make files public, send them out for review, collect files from anyone, and save files from links. Nothing gets deleted.
Everythingfiles:fullWhat it can doAlso delete files.

One folder#

Held to a folder, an app sees that folder and the ones inside it. Files it adds land there, and the tools that reach across the whole drive, like collections and image generation, are left out.

How long#

Choose when it stops working: never, a day, a week, a month or 3 months. Until then its access token lasts an hour at a time and renews itself.

A connection nobody uses for six months stops on its own. You can connect the same app more than once, say from two Claude accounts, and keep up to 10 connections to it. Past that, the one used least gives way.

Uploading#

Ask Claude to save an attachment or a file it made, and it lands in your drive. upload_file takes the file's name and size, Claude sends the bytes straight to storage, and complete_upload checks them and saves the file. It arrives private, in the connection's folder, or one you name.

It needs Organize access. Big files go in parts: Claude calls upload_file again with the same ticket for the next ones, or to pick an interrupted upload back up. A ticket lasts a day, and nothing is saved until complete_upload succeeds.

On your machine#

freaking-fast-mcp runs the drive's tools on your machine. Its upload_file uploads straight from your disk in one step, and it makes images. It needs Node 20 or later and an API token with full access.

Make an API token
Claude Code
claude mcp add freakingfast -e FREAKING_FAST_API_TOKEN=<token> -- npx -y freaking-fast-mcp
Any client that reads mcpServers
{
  "mcpServers": {
    "freakingfast": {
      "command": "npx",
      "args": ["-y", "freaking-fast-mcp"],
      "env": { "FREAKING_FAST_API_TOKEN": "<token>" }
    }
  }
}
  • Uploads stream from disk, so a 2 TB file never has to fit in memory.
  • Storage is checked before an upload starts. If a file won't fit, you find out right away.
  • Uploading the same file again, with the same name, size and modified time, returns the one already stored.
  • The token buys a session that lasts a day, renewed a minute before it runs out.

Sharing with someone's AI#

Share a folder or one file with someone else's AI, from the app or by asking your own: “share the Harbor folder with Sam's Claude”, or “share the interview with Ana's ChatGPT”. You get a link. Whoever you send it to adds it to their AI as a connector, with no account. For a folder, it can find, read and download what's in it and the folders inside it; for a file, it can read it, search inside it and download it. It changes nothing, and sees nothing else in your drive.

A share link
https://mcp.freakingfast.io/s/abs_…

In Claude it's a custom connector, in ChatGPT a connector in developer mode with no authentication, and in Claude Code claude mcp add --transport http harbor <link>. Opened in a browser, the link says the same. It stops on the day you chose, within a year, and you can stop it sooner from Agents, where you can also copy it again. A share stops when its folder or file is deleted, and a shared file in the trash opens nothing until it's back. What their AI does through it is listed on Agents with everything else.

To get files back, send a file request's link. An AI connected to the drive, or to one of your shares, sends files to it with deliver_file and complete_delivery, the bytes going from its own environment straight to storage. An AI with no connection at all adds the request's own connector address, which the request page shows:

A request, for an AI
https://mcp.freakingfast.io/r/<request id>

Tools#

Every tool, the least it needs, and whether an app held to one folder still has it. The connector has all of them but the two image tools, which only the server on your machine has. Its upload_file takes a name and size rather than a path.

Files#

ToolWhat it doesNeeds
upload_fileUploads a file: from disk on your machine, or sent by the app through the connectorNeedsOrganizeConnector: the app sends the bytes
complete_uploadChecks an upload sent through the connector and saves the file, privateNeedsOrganizeConnector only
import_from_urlBrings a file at a link into the drive, fetched by the driveNeedsShare
get_download_urlGets a fresh 6-hour link for a stored file, or a public file's lasting pageNeedsRead
read_fileShows the model what's in a file: an image, a video's frame, a PDF's or text file's text, what's saidNeedsRead
set_file_visibilityMakes files public or privateNeedsOrganizeShare to make public
list_filesLists the library, newest first, a page at a timeNeedsRead
search_filesFinds files by words, type, tag, folder, date, length, size, tempo, key or place, or what's inside themNeedsRead
delete_filesMoves files to the trash, or deletes them for goodNeedsEverything
restore_filesBrings files back from the trashNeedsOrganize
get_storage_usageReports used, limit, and remaining storageNeedsReadWhole drive

Organize#

ToolWhat it doesNeeds
tag_filesAdds or removes tags by nameNeedsOrganize
move_filesPuts files in a folder by path, making it if neededNeedsOrganize
list_collectionsLists folders, playlists, galleries and tagsNeedsReadWhole drive
create_playlistMakes a playlist from audio and videoNeedsOrganizeWhole drive
add_to_playlistAdds audio and video to a playlistNeedsOrganizeWhole drive
create_galleryMakes a photo gallery and picks what goes up frontNeedsOrganizeWhole drive
add_to_galleryAdds images to a galleryNeedsOrganizeWhole drive
set_coverPuts an image on a track, video or playlist as its coverNeedsOrganizeWhole drive

Collect and review#

ToolWhat it doesNeeds
create_file_requestMakes a link anyone can send files to, no account neededNeedsShare
list_file_requestsShows what arrived on each request, and from whomNeedsRead
close_file_requestStops a request taking files, or opens it againNeedsOrganizeShare to reopen
create_review_linkSends a video or track out for timestamped notesNeedsShare
get_review_notesReads the notes as an edit list, by timecodeNeedsRead
reply_to_noteAnswers a note, like "fixed in v3"NeedsShareWhole drive
resolve_noteMarks a note dealt with, or opens it againNeedsOrganizeWhole drive

Keep up#

ToolWhat it doesNeeds
get_file_viewsReports how often files were opened, and which people they were shared with didNeedsRead
get_changesCatches up on what happened: files sent in, notes, views, new files, finished images and jobsNeedsRead

Work on files#

ToolWhat it doesNeeds
run_jobCuts, converts, zips and more, where the files are stored, beside themNeedsOrganize
get_jobChecks on a job, and gets what it madeNeedsRead
stop_jobStops a job, keeping what it already madeNeedsOrganize

Standing orders#

ToolWhat it doesNeeds
create_standing_orderPlans something the drive does by itself each time files arrive or notes come in, for you to startNeedsShareConnector only
list_standing_ordersShows the standing orders, and how each one's last run wentNeedsReadWhole driveConnector only

Other people's AI#

ToolWhat it doesNeeds
share_with_agentShares a folder or a file with someone else's AI by a link, read only, until a dateNeedsShare
get_file_requestSays what a file request link asks for, and whether it still takes filesNeedsReadConnector only
deliver_fileSends a file from the AI's own environment to a file request linkNeedsReadConnector only
complete_deliveryChecks a file sent to a request and files it; the last one tells whoever askedNeedsReadConnector only

Images#

ToolWhat it doesNeeds
generate_imageMakes an image with the plan's models and saves it to the driveNeedsOrganizeWhole driveNot on the connector
get_generated_imagePicks up an image that was still being madeNeedsReadWhole driveNot on the connector

Activity, undo and disconnecting#

Every app you connect, every folder you share, and every tool they call, is listed under Agents in the app, newest first, with what it touched. Disconnecting an app there stops it within a minute, and ends anything it had waiting. Stopping a share stops its link at once.

Undo takes back one thing an app did, or everything it did: files it trashed come back, and moves, tags, sharing and new uploads go back to how they were. Anything someone changed again since is left as it is. Through the connector, an app can only move files to the trash, so a delete can always be undone for 30 days.