Skip to content

MCP server

Let your AI assistant convert files

One remote server, five tools, 584 conversions. Claude, Cursor, VS Code, Windsurf and ChatGPT can check whether a pair is supported, convert a file from bytes or a link, and read the result inline. Setup is one line for Claude Code and a short JSON block for the others.

Create a token HTTP API docs
5
tools
65
formats
HTTP
Streamable transport
Starter+
plan required

Setup

Connect in two steps

  1. 1Create a token. Sign in, open Dashboard → MCP Server, name the connection after the client (“Cursor on laptop”) and copy the token; it is shown once. Starter plan or higher.
  2. 2Paste the snippet for your client below, replace the token, restart the client, then ask it: “check whether you can convert a HEIC to JPG, then list the formats you support”.

Where: Run once in your terminal

bash
claude mcp add --transport http filetypeconverters https://filetypeconverters.com/api/mcp \
  --header "Authorization: Bearer ftc_mcp_YOUR_TOKEN"

ftc_mcp_YOUR_TOKEN is a placeholder — create a credential on the right and it's filled in for you.

Then ask Claude: “list the formats you can convert” to confirm it's live.

Endpoint https://filetypeconverters.com/api/mcp · protocol 2025-06-18 · JSON responses (no SSE stream) · header Authorization: Bearer ftc_mcp_…

Tools

What the assistant can call

  • list_formatsThe 65 formats with ids, names, extensions and family. Use the ids as from/to elsewhere.
  • check_conversionIs A → B supported? Returns the engine, whether AI Deep Read is required, the options the engine accepts, and alternatives when the pair is not supported. Accepts ids, extensions (jpeg) or plain names (word, excel).
  • list_conversionsEvery target for a source format, or every source that can produce a target. Compact rows: from, to, slug, engine, mode.
  • convert_fileConvert base64 content or a public https URL. Text outputs up to 200 KB come back inline so the assistant can read them; binaries come back as base64 plus a download URL that accepts the same token. deepRead: true reads scans and photos (1 credit per page).
  • get_conversionFetch a previous conversion by id: status, warnings and the output if it is still retained. Poll this for queued jobs.

A real exchange

Recorded against this server. The assistant checks the pair first, then converts a CSV by URL and receives the JSON inline.

text
→ tools/call check_conversion { "from": "pdf", "to": "word" }
← { "supported": true, "from": "pdf", "to": "docx", "engine": "vision",
    "engineLabel": "AI Deep Read", "mode": "ai_deep_read", "options": [] }

→ tools/call convert_file { "url": "https://example.com/q3.csv", "to": "json" }
← { "id": "280d0f31-…", "status": "succeeded", "from": "csv", "to": "json",
    "bytes": 614, "warnings": [],
    "text": "[\n  { \"company\": \"Northwind Traders\", \"arr\": 48000, … } ]",
    "downloadUrl": "https://filetypeconverters.com/api/v1/jobs/280d0f31-…/download" }

In practice

Prompts that work

  • “Convert every .heic in ./photos to JPG and save them next to the originals.”

    Claude Code reads each file, calls convert_file, and writes the base64 result to disk.

  • “Here is a bank statement PDF — give me the transactions as a spreadsheet.”

    check_conversion reports pdf → xlsx needs AI Deep Read; convert_file with deepRead: true returns the workbook; one credit per page.

  • “Pull https://…/report.docx and summarise it.”

    convert_file with the URL and to: md returns Markdown inline, which the assistant reads directly.

  • “What formats can you turn a Kindle AZW3 into?”

    list_conversions with from: azw3 lists EPUB, PDF, MOBI, DOCX, Markdown and text.

Questions about the MCP server

What is MCP?

The Model Context Protocol is an open standard (modelcontextprotocol.io) that lets an AI application call tools exposed by a server. Claude Code, Claude Desktop, Cursor, Windsurf, VS Code and ChatGPT all speak it. This server exposes file conversion as five tools over Streamable HTTP, so the assistant can check a pair, convert a file and read the result without leaving the chat.

How does the assistant get the file to you?

Two ways. It can read the file and send the bytes as base64 in the tool call, which suits files up to a few megabytes. Or it can pass a public https URL and the server fetches it (private addresses are refused). Outputs that are text come back inline; anything else comes back as base64 and as a download URL the assistant or you can fetch with the same token.

Does it cost extra?

No. MCP access is included on Starter and above; conversions made through it count against the same monthly allowance as the web app and API (Starter: 500 a month). AI Deep Read reads use credits at one per page, exactly as in the browser.

Is the token safe to put in a config file?

A token is scoped to one workspace with convert and read permissions only: it cannot change billing, invite members or read other workspaces. Create one token per client so you can revoke a single laptop or agent from Dashboard → MCP Server without touching the others. Tokens are shown once at creation and stored hashed.

Why does the client say it connected but list no tools?

Almost always a plan or token issue: the workspace is on Free (MCP needs Starter or higher) or the token was revoked. The dashboard has a Test connection button that runs tools/list with your token and shows the exact error. The server also answers ping, resources/list and prompts/list with empty results so clients that probe them do not log noise.

Can I use the same token for the HTTP API?

Yes for reads and conversions. Tokens that start with ftc_mcp_ are accepted by /api/v1/convert, /api/v1/jobs and the download endpoints with the same permissions. For batch creation or usage reporting create an API key instead.