AI tools and MCP

The main SumOffice MCP server works with your file: the agent opens a document or workbook in a private editor cabin, reads it the way the engine does, proposes an edit and applies it after you say yes. It is described in section 1.

The second server is about the documentation. The reference is large — 41 session commands, 321 SumDoc shell methods, 487 request and response types and 208 SumSheet contract operations — and neither a person nor an assistant needs to keep it in their head. That is why the site offers its content in three machine-readable forms: section 2, llms.txt and the widget on every page.

1. Product MCP server

This server works not with the documentation but with your document or workbook. The agent opens the file in a private editor cabin on the viewing server, reads it the way the engine does — formulas, macros, Power Query, paragraphs, comments — proposes an edit, applies it after you say yes and hands back a copy that Word and Excel open without a repair dialog. The original is never changed.

Connect in one minute

Address: https://view.sumoffice.com/mcp, streamable HTTP. Any MCP client works; no account, no key. Each session gets its own cabin; files and copies are deleted after 24 hours.

claude_desktop_config.json

{ "mcpServers": {
    "office": {
      "command": "npx",
      "args": ["-y", "mcp-remote", "https://view.sumoffice.com/mcp"]
    }
} }

Claude Code:

claude mcp add --transport http office https://view.sumoffice.com/mcp

Cursor and other clients connect the same URL as a remote server.

Local files. The same server is available as the Docker image hissih/sumoffice-mcp (linux/amd64 and arm64). It runs on your side, accepts a path inside the mounted folder, and opens the file itself in a sandbox on the server, just like the remote variant:

{
  "mcpServers": {
    "office-agent": {
      "command": "docker",
      "args": ["run", "-i", "--rm", "-v", "/path/to/files:/work",
               "hissih/sumoffice-mcp"]
    }
  }
}

Then — document_open_file with path: "/work/report.xlsm".

Step	Tools
Open	document_open_file — your file by url, as base64 with a name or, in the Docker image, by path: .docx, .doc, .odt, .xlsx, .xlsm, .xlsb, .xls, .ods, up to 50 MB. document_open — a demo file by name, to try it out.
Read a workbook	sheet_describe (sheets, used ranges, revision), sheet_read (values and formulas of a range), sheet_find; macro_list, macro_read_source, macro_plan — the VBA project: entry points, routes, obstacles, nothing is run; query_list, query_explain — Power Query steps and where they load to; workbook_inspect — the whole workbook model in one call.
Read a document	document_describe, document_read (paragraphs in a window), document_find, document_comments (comments and tracked changes).
Change	sheet_propose_edits → sheet_apply: up to 20 cells, the engine first shows what will change, whether errors will appear and whether input numbers get overwritten. document_propose_replacement → document_apply: new text for one paragraph, with “before — after” and checks. Nothing changes before the apply call; a proposal against an outdated revision is rejected, not applied silently.
Collect	document_save_copy — the changed file as a link valid for 24 hours; a workbook with macros comes back as .xlsm with the VBA project untouched. document_export_pdf — the document as laid out by the engine, as a PDF with the page count.

The same cycle — look, propose, wait for permission, apply — is available in the agent panel inside the editor; how it works and why the refusal comes before the change — VBA, Power Query and agents.

Boundaries set on purpose

One file per session; the session closes after 10 minutes of silence; up to five simultaneous sessions per server. Cell edits are values and formulas in a rectangular window; a paragraph edit replaces its text and keeps the formatting. The server does not write to your original, does not run macros and does not refresh queries: the agent reads and plans them, and running them on a copy so far lives only in the editor’s agent panel. Need the same server inside your network, next to your files — contact us.

2. Documentation MCP server

The documentation site runs its own MCP server at /mcp — not to be confused with the product server above: this one answers questions about the documentation and does not work with your file. Tools:

Tool	Arguments	What it returns
search_docs	query	Up to five most relevant pages with URL, description and an excerpt around the match.
get_doc	path — page identifier or URL	The full text of the page in markdown.
list_docs	—	All pages with titles and descriptions.

The tool catalog in machine-readable form: /mcp-schema.json.

Claude Code

claude mcp add --transport http fastoffice-docs https://docs.sumoffice.com/mcp

Cursor

.cursor/mcp.json

{
  "mcpServers": {
    "fastoffice-docs": { "url": "https://docs.sumoffice.com/mcp" }
  }
}

VS Code

.vscode/mcp.json

{
  "servers": {
    "fastoffice-docs": {
      "type": "http",
      "url": "https://docs.sumoffice.com/mcp"
    }
  }
}

Clients that only speak stdio

{
  "mcpServers": {
    "fastoffice-docs": {
      "command": "npx",
      "args": ["mcp-remote", "https://docs.sumoffice.com/mcp"]
    }
  }
}

Why this matters

An assistant connected to the MCP server stops guessing command names and request fields. Instead of an “approximately like this” call to apply_cell_edits, it gets the actual field table from the cells reference — together with a link to the source page.

3. llms.txt

For tools that work with text dumps, the site publishes six files. All of them are generated at build time from the same content as the site and open by direct link.

File	What is inside
/llms.txt	The table of contents: links to the other files with descriptions. The entry point.
/llms-full.txt	The whole documentation in one file, about 2.9 MB.
/llms-small.txt	The same material without the reference field tables, about 2.2 MB. That is still a lot: if the model has a small context, take a topical set below instead.
/_llms-txt/a4-reference.txt	The SumDoc reference only: Session Protocol, window.fastdocShell, DTOs, events.
/_llms-txt/f1-reference.txt	The SumSheet reference only: Compute Contract, 208 operations.
/_llms-txt/platform-overview.txt	The product part: white-label delivery, security, the pilot.

Why topical sets

The whole documentation takes about three megabytes — a noticeable share of a context window. If the question is only about SumSheet operations, it is cheaper and more precise to load the single f1-reference.txt (about 820 KB) than the whole documentation.

4. The chat assistant on the site

The button in the bottom-right corner opens a chat about the documentation. It is a full conversation: you can refine as you go without repeating the context. The answer streams in, and the conversation is kept in the browser until your next visit.

Under every answer is a list of sources. Check them: the answer is worded by a model, and it may be inaccurate.

Answers are built from the documentation only

The assistant answers solely on the basis of the pages of this site. If the documentation does not have what you need, it says so instead of inventing a command or field name.

If answers are switched off — no model key is set or the quota has run out — the chat shows the sections it found, with links.

5. The “Open in…” buttons on pages

The header of every documentation page has an actions menu: copy the page as markdown or open its content in ChatGPT or Claude. This lets you ask about one specific page without connecting MCP.

What to choose

Task	Tool
Writing integration code in an IDE with an assistant	The MCP server for the documentation
The agent has to read or fix my file	The MCP server for the product
A one-off question about the open page	The “Open in…” button
Quickly find where something is described	The “Ask” widget or site search
Loading the documentation into my own pipeline	/llms-full.txt or /mcp/docs-index.json