SumOffice — real Word and Excel for agents

Official21 tools

by SumOfficeApp

Real Word and Excel for agents: open your .docx/.xlsm, read, propose, apply, get it back intact.

Model Context Protocol (MCP) Server: com.sumoffice/office

The com.sumoffice/office MCP server is designed for working with real Word and Excel files for agents. It supports opening .docx and .xlsm, reading their contents, proposing changes, and applying updates. The server returns documents intact after processing.

🛠️ Key Features

Handles .docx and .xlsm
Read document contents
Propose changes
Apply updates
Returns files intact

🚀 Use Cases

Agent workflows that need to update Word and Excel documents
Reading existing documents, making agent-generated proposals, and applying them

⚡ Developer Benefits

Maintains original document integrity (“get it back intact”)
Streamlines document edit cycles for agents across Word and Excel formats

⚠️ Limitations

Available details do not specify support beyond .docx and .xlsm

Captured live from the server via tools/list.

document_open

Open one of the demo files by name: documents footnotes, comments, forms, rtl, wild-horse, workshop, offer, tools; workbooks workbook, functions, macros, queries, gep, vba-web, sqlite, order-desk, onrr, cashflow, pivot, chart, notes, check, xtools. Default is footnotes. For your own file use document_open_file. Opens a separate editor cabin for this session.

Parameters1

slug

string

required

Raw schema

{
  "type": "object",
  "properties": {
    "slug": {
      "type": "string",
      "enum": [
        "footnotes",
        "comments",
        "forms",
        "rtl",
        "wild-horse",
        "workshop",
        "workbook",
        "functions",
        "macros",
        "queries",
        "gep",
        "vba-web",
        "sqlite",
        "order-desk",
        "onrr",
        "cashflow",
        "offer",
        "pivot",
        "chart",
        "notes",
        "check",
        "tools",
        "xtools"
      ]
    }
  },
  "required": [
    "slug"
  ],
  "$schema": "http://json-schema.org/draft-07/schema#"
}

document_open_file

Open YOUR file in a private editor cabin for this session: a .docx/.doc/.odt document or an .xlsx/.xlsm/.xlsb/.xls/.ods workbook, up to 50 MB. Give exactly one of: url (https link the server can fetch), base64 (file bytes; then name with the extension is required). Nothing in the file is executed. The file is kept for 24 hours and then deleted. After opening, use document_* tools for a document or sheet_* / macro_* / query_* / workbook_inspect for a workbook; sheet_propose_edits → sheet_apply and document_propose_replacement → document_apply change the copy in the cabin; document_save_copy returns a download link of the edited file.

Parameters4

url	string	optional
base64	string	optional
path	string	optional
name	string	optional

Raw schema

{
  "type": "object",
  "properties": {
    "url": {
      "type": "string",
      "maxLength": 2000
    },
    "base64": {
      "type": "string"
    },
    "path": {
      "type": "string",
      "maxLength": 1000
    },
    "name": {
      "type": "string",
      "maxLength": 200
    }
  },
  "$schema": "http://json-schema.org/draft-07/schema#"
}

document_describe

Name, pages, paragraphs, words, number of comments and tracked changes, whether track changes is on. For documents; call first.

Parameters

No parameters.

Raw schema

{
  "type": "object",
  "properties": {},
  "$schema": "http://json-schema.org/draft-07/schema#"
}

document_read

Paragraph text as a window: from = index of the first paragraph (0-based), count = how many (max 40). The answer carries next — where to continue, null at the end. A missing paragraph is a refusal, not an invention.

Parameters2

from	integer	required
count	integer	optional

Raw schema

{
  "type": "object",
  "properties": {
    "from": {
      "type": "integer",
      "minimum": 0,
      "maximum": 9007199254740991
    },
    "count": {
      "type": "integer",
      "minimum": 1,
      "maximum": 40
    }
  },
  "required": [
    "from"
  ],
  "$schema": "http://json-schema.org/draft-07/schema#"
}

document_find

Paragraphs containing the text (case-insensitive), with snippets around the match.

Parameters1

query

string

required

Raw schema

{
  "type": "object",
  "properties": {
    "query": {
      "type": "string",
      "minLength": 2,
      "maxLength": 200
    }
  },
  "required": [
    "query"
  ],
  "$schema": "http://json-schema.org/draft-07/schema#"
}

document_comments

Comments (author, text, anchor, resolved) and tracked changes — up to 50.

Parameters

No parameters.

Raw schema

{
  "type": "object",
  "properties": {},
  "$schema": "http://json-schema.org/draft-07/schema#"
}

document_propose_replacement

Propose the full new text of one paragraph (paragraph_index from document_read). Nothing changes yet: the editor core builds the change on a copy and returns before/after plus checks (numbers, dates, links kept). Then call document_apply with the proposal_id. Read the paragraph first.

Parameters3

paragraph_index	integer	required
text	string	required
reason	string	optional

Raw schema

{
  "type": "object",
  "properties": {
    "paragraph_index": {
      "type": "integer",
      "minimum": 0,
      "maximum": 9007199254740991
    },
    "text": {
      "type": "string",
      "minLength": 1,
      "maxLength": 20000
    },
    "reason": {
      "type": "string",
      "maxLength": 300
    }
  },
  "required": [
    "paragraph_index",
    "text"
  ],
  "$schema": "http://json-schema.org/draft-07/schema#"
}

document_apply

Apply a proposal from document_propose_replacement to the document in this session's cabin (not to your original file). Stale proposal (the document moved on) is refused by the engine — propose again. Repeating an applied proposal is a no-op.

Parameters1

proposal_id

string

required

Raw schema

{
  "type": "object",
  "properties": {
    "proposal_id": {
      "type": "string",
      "minLength": 4,
      "maxLength": 64
    }
  },
  "required": [
    "proposal_id"
  ],
  "$schema": "http://json-schema.org/draft-07/schema#"
}

sheet_describe

Workbook overview: every sheet tab with its used range, the current revision, and the header row of the sheet that is open. Call this first, then pass the tab name you need as "sheet" to sheet_read and sheet_find.

Parameters

No parameters.

Raw schema

{
  "type": "object",
  "properties": {},
  "$schema": "http://json-schema.org/draft-07/schema#"
}

sheet_read

Read a rectangular range like A1:F20 (max 60 rows × 26 columns): values and formulas. Empty cells are omitted. ALWAYS pass "sheet" with the exact tab name when the workbook has more than one sheet — without it you read whichever sheet happens to be open, and the answer will quietly be about the wrong sheet. Call sheet_describe first if you do not know the tab names.

Parameters2

range	string	required
sheet	string	optional	Exact tab name, e.g. "Logical". Required when the workbook has more than one sheet.

Raw schema

{
  "type": "object",
  "properties": {
    "range": {
      "type": "string"
    },
    "sheet": {
      "type": "string",
      "description": "Exact tab name, e.g. \"Logical\". Required when the workbook has more than one sheet."
    }
  },
  "required": [
    "range"
  ],
  "$schema": "http://json-schema.org/draft-07/schema#"
}

sheet_find

Find cells whose value or formula contains the text (case-insensitive) in the used range of a sheet. ALWAYS pass "sheet" with the exact tab name when the workbook has more than one sheet: without it only the sheet that happens to be open is searched.

Parameters2

query	string	required
sheet	string	optional	Exact tab name. Required when the workbook has more than one sheet.

Raw schema

{
  "type": "object",
  "properties": {
    "query": {
      "type": "string",
      "minLength": 1,
      "maxLength": 100
    },
    "sheet": {
      "type": "string",
      "description": "Exact tab name. Required when the workbook has more than one sheet."
    }
  },
  "required": [
    "query"
  ],
  "$schema": "http://json-schema.org/draft-07/schema#"
}

macro_list

List VBA modules and runnable macros (entry points) in this workbook with their route and blockers. Runs nothing.

Parameters

No parameters.

Raw schema

{
  "type": "object",
  "properties": {},
  "$schema": "http://json-schema.org/draft-07/schema#"
}

macro_read_source

Read the source text of one procedure (module + procedure name from macro_list). Read-only; runs nothing.

Parameters2

module	string	required
procedure	string	required

Raw schema

{
  "type": "object",
  "properties": {
    "module": {
      "type": "string"
    },
    "procedure": {
      "type": "string"
    }
  },
  "required": [
    "module",
    "procedure"
  ],
  "$schema": "http://json-schema.org/draft-07/schema#"
}

macro_plan

Plan a macro run without running it: route, whether the engine can run it natively, blockers, instruction budget.

Parameters2

entrypoint	string	required	Module.Procedure
active_sheet	string	optional

Raw schema

{
  "type": "object",
  "properties": {
    "entrypoint": {
      "type": "string",
      "description": "Module.Procedure"
    },
    "active_sheet": {
      "type": "string"
    }
  },
  "required": [
    "entrypoint"
  ],
  "$schema": "http://json-schema.org/draft-07/schema#"
}

query_list

List Power Query queries in this workbook: names, step chain, load target, connections. Reads only; refreshes nothing.

Parameters

No parameters.

Raw schema

{
  "type": "object",
  "properties": {},
  "$schema": "http://json-schema.org/draft-07/schema#"
}

query_explain

Explain one Power Query query by name: its steps in order, where it loads, its connection. Reads only; refreshes nothing.

Parameters1

name

string

required

Raw schema

{
  "type": "object",
  "properties": {
    "name": {
      "type": "string"
    }
  },
  "required": [
    "name"
  ],
  "$schema": "http://json-schema.org/draft-07/schema#"
}

workbook_inspect

Map the whole application model of this workbook in one call: sheets; VBA modules, procedures, which can run here and which are blocked and why; event handlers; external objects (CreateObject) and Windows API (Declare) dependencies; project references; Power Query queries; data connections; form controls, ActiveX, embedded objects (preserved). Runs nothing. Call this first for questions like 'how is this workbook built' or 'what does it depend on'.

Parameters

No parameters.

Raw schema

{
  "type": "object",
  "properties": {},
  "$schema": "http://json-schema.org/draft-07/schema#"
}

sheet_propose_edits

Propose new contents for up to 20 cells (formulas like =SUM(B2:B9), or numbers/text). Nothing changes in the workbook: the engine previews the edits on the current revision and reports what would change and whether errors appear; the person decides whether to apply. Read the cells first. Do not overwrite input numbers unless the task says so. Then call sheet_apply with the proposal_id.

Parameters3

edits	array	required
sheet	string	optional
reason	string	optional

Raw schema

{
  "type": "object",
  "properties": {
    "edits": {
      "minItems": 1,
      "maxItems": 20,
      "type": "array",
      "items": {
        "type": "object",
        "properties": {
          "cell": {
            "type": "string"
          },
          "input": {
            "type": "string",
            "maxLength": 2000
          }
        },
        "required": [
          "cell",
          "input"
        ]
      }
    },
    "sheet": {
      "type": "string"
    },
    "reason": {
      "type": "string",
      "maxLength": 300
    }
  },
  "required": [
    "edits"
  ],
  "$schema": "http://json-schema.org/draft-07/schema#"
}

sheet_apply

Apply a proposal from sheet_propose_edits to the workbook in this session's cabin (not to your original file); the engine recalculates. Refused if the workbook revision moved on since the proposal — propose again. Repeating an applied proposal is a no-op.

Parameters1

proposal_id

string

required

Raw schema

{
  "type": "object",
  "properties": {
    "proposal_id": {
      "type": "string",
      "minLength": 4,
      "maxLength": 64
    }
  },
  "required": [
    "proposal_id"
  ],
  "$schema": "http://json-schema.org/draft-07/schema#"
}

document_save_copy

Write the document or workbook as it is now in this session's cabin (after applied proposals) to a copy and return a download link valid for 24 hours. Your original file is never touched. A workbook with macros comes back as .xlsm with the VBA project intact; a document keeps its styles, comments and tracked changes.

Parameters1

inline

boolean

optional

also return the bytes as base64 (only for files up to 4 MB)

Raw schema

{
  "type": "object",
  "properties": {
    "inline": {
      "description": "also return the bytes as base64 (only for files up to 4 MB)",
      "type": "boolean"
    }
  },
  "$schema": "http://json-schema.org/draft-07/schema#"
}

document_export_pdf

Lay out the document as it is now in this session's cabin (after applied proposals, or as opened) and return a PDF download link valid for 24 hours, plus the page count. Documents only; for workbooks use document_save_copy.

Parameters

No parameters.

Raw schema

{
  "type": "object",
  "properties": {},
  "$schema": "http://json-schema.org/draft-07/schema#"
}

README.md3 KBView on GitHub

SumOffice MCP server — real Word and Excel for agents

One server, two ways to run it. The agent opens your .docx / .xlsx / .xlsm, reads it, proposes a change, you approve, the file comes back intact — Excel and Word open the result without a repair dialog. Macros (VBA) and Power Query travel with the file.

Registry name: com.sumoffice/office · page with proof: https://sumoffice.com/mcp · image: docker.io/hissih/sumoffice-mcp

Remote — nothing to install

https://view.sumoffice.com/mcp (streamable HTTP).

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

Clients without HTTP support: npx -y mcp-remote https://view.sumoffice.com/mcp (see examples/).

Local over stdio — local file paths

docker run -i --rm -v "$PWD:/work" hissih/sumoffice-mcp

The server runs on your machine and accepts path in document_open_file (paths under /work). The file itself is opened in a private, short-lived editor session on the SumOffice server and deleted after the session; nothing in the file is executed.

Tools

Documents (.docx)	Spreadsheets (.xlsx/.xlsm/.xlsb)	Both
`document_open_file`, `document_open`, `document_describe`, `document_read`, `document_find`, `document_comments`, `document_propose_replacement` → `document_apply`	`sheet_describe`, `sheet_read`, `sheet_find`, `sheet_propose_edits` → `sheet_apply`, `macro_list`, `macro_read_source`, `macro_plan`, `query_list`, `query_explain`, `workbook_inspect`	`document_save_copy`, `document_export_pdf`

propose_* never writes: it returns a diff for the person or the client to approve; *_apply applies exactly that proposal. Every step is listed in the response, so the agent cannot change a file silently.

Examples

examples/claude-desktop.json — Claude Desktop (claude_desktop_config.json), local stdio via Docker.
examples/cursor-mcp.json — Cursor (.cursor/mcp.json), remote HTTP.
examples/claude-code.sh — Claude Code, both variants.
examples/docker-compose.yml — the local server as a compose service with a mounted folder.

What it does not do yet

One editor session per document at a time; a second agent on the same file gets read-only.
Power Query definitions are preserved, not refreshed, in the server session.
Macros are read and routed (which will run natively, which stay in Excel); they are not executed by the MCP server.

Limits and privacy

Free, no account. Remote: per-address daily quota; files are deleted after the session or 24 hours at most, never used for anything else. Details: https://sumoffice.com/app-privacy.html

Licence

This repository (documentation, manifest, examples) is MIT. The server image is free for evaluation and for use with the public endpoint; production use inside an organisation is licensed per application — https://sumoffice.com/download#server. Questions: hello@sumoffice.com

Install

Tools (21)

document_opendocument_open_filedocument_describedocument_readdocument_finddocument_commentsdocument_propose_replacementdocument_apply

Configuration

Arguments

-vrequireddefault {files_dir}:/work

Add to VS Code Add to Cursor

claude_desktop_config.json

json

{
  "mcpServers": {
    "office": {
      "command": "docker",
      "args": [
        "run",
        "-i",
        "--rm",
        "docker.io/hissih/sumoffice-mcp:0.3.1"
      ]
    }
  }
}

Model Context Protocol (MCP) Server: com.sumoffice/office

🛠️ Key Features

🚀 Use Cases

⚡ Developer Benefits

⚠️ Limitations

document_open

document_open_file

document_describe

document_read

document_find

document_comments

document_propose_replacement

document_apply

sheet_describe

sheet_read

sheet_find

macro_list

macro_read_source

macro_plan

query_list

query_explain

workbook_inspect

sheet_propose_edits

sheet_apply

document_save_copy

document_export_pdf

SumOffice MCP server — real Word and Excel for agents

Remote — nothing to install

Local over stdio — local file paths

Tools

Examples

What it does not do yet

Limits and privacy

Licence

Install

Configuration

Related servers