MCP server: use Documentize inside Claude, Cursor, VS Code
Open the app: Connect the MCP server → (products.documentize.ai)
Summary
The Documentize MCP server lets an AI assistant call the same document operations as the web apps. You say "compress this and give me a summary"; the assistant calls tools, waits for the tasks and hands you download links. Page: products.documentize.ai/pdf-mcp.
Endpoint and identity
- Streamable HTTP endpoint:
https://api.documentize.app/mcp. The server is stateful: it returns anMcp-Session-Idheader afterinitialize; MCP clients handle this for you. - Anonymous works with no setup, under a tiny daily allowance (documented as 2 operations in the repo; verify current value).
- Registered users get a larger allowance (5 per day for free accounts, documented; verify), subscribers the paid allowance, the same as on the website.
- To be recognized, create a long-lived key on the MCP page (it starts with
dmk_, is shown once, and is stored only as a hash). Send it asAuthorization: Bearer dmk_.... - Clients that implement MCP authorization can use OAuth instead of a key.
One operation is one processing call, whatever the file count. Status polling and account tools are free.
Client configuration
Claude Desktop, claude_desktop_config.json:
{
"mcpServers": {
"documentize": {
"type": "http",
"url": "https://api.documentize.app/mcp",
"headers": { "Authorization": "Bearer dmk_your_key_here" }
}
}
}
VS Code (GitHub Copilot, Agent mode), .vscode/mcp.json:
{
"servers": {
"documentize": {
"type": "http",
"url": "https://api.documentize.app/mcp",
"headers": { "Authorization": "Bearer dmk_your_key_here" }
}
}
}
Restart the client; the tools appear in its tool list. Drop the headers block to connect anonymously. Cursor and other clients use their own MCP settings with the same URL and header (see the MCP page).
The tools
Names as registered by the server (tools.json):
- Convert and organize: ConvertDocument, MergeDocuments, SplitDocument, CompressDocument, RotateDocument, ResizeDocument, RemovePages
- Security: LockDocument, UnlockDocument, SignDocument, VerifySignature
- Extract and search: ExtractText, ExtractImages, ExtractMetadata, ExtractFormData, ParseDocument, SearchInDocument
- OCR and finishing: MakeSearchable, RunOcr, FlattenDocument, AddTable, AddToc, AddIllustrations
- AI: GenerateTableOfContents, GenerateAbstract, GenerateChecklist, AnalyzeResume, ChatWithDocument, GenerateSvg
- Batch: ProcessBatch, WaitForTasks, ListBatchOperations
- Account: GetAccountStatus, GetSignInInstructions, SignOut
- Flow: GetTaskStatus, GetNextAppSuggestions
Every processing tool returns a task id at once; the result arrives later, via GetTaskStatus (one task) or WaitForTasks (many).
An example conversation
You: Here are three scans (URLs). Make them one searchable PDF, drop the first page, and shrink it.
Assistant: calls MergeDocuments with the three URLs, then GetTaskStatus until done. Merged. Calls MakeSearchable on the result URL. Calls RemovePages (page 1), then CompressDocument. Here is the final download link. It saved about N MB (the assistant reports the real figure).
The pattern: each step's download link becomes the next step's input. For 20 or more files, the assistant should use ProcessBatch (one operation applied to up to 200 files per call, parallel on the server) followed by WaitForTasks, which returns every link in one call. A wait is capped at a few minutes; call it again with the ids still pending.
When something is refused
Refusals are written for a model: a stable error code and a next step. quota_exceeded means out for today; credential_rejected means the key is revoked or mistyped (replace it, do not retry); auth_required means the anonymous allowance is spent. Ask the assistant to run GetAccountStatus to see tier and remaining operations.
Cautions
- Files travel to the cloud; use self-host for confidential material.
- Exact tool parameters are discovered by the client from the server; do not guess them, ask the assistant to list them (ListBatchOperations for the batch ones).