OfficeAgent.NET

Give coding agents a structured way to create and edit real Word documents and
PowerPoint decks. OfficeAgent.NET turns an agent's intent into typed, validated
operations and applies them directly to OOXML packages while preserving document
structure.

Use it to generate documents and presentations, make targeted edits, update tables,
styles, and images, or manage comments and review state. The engine ships as an MCP
server, Microsoft Agent Framework tools, and a .NET API, with filesystem, SharePoint,
session, and inline document workflows.

One example is a targeted Word edit whose result remains reviewable:

What this project does

A .docx or .pptx file is a package of related XML parts. A small text change
can affect runs, styles, numbering, comments, content controls, or revision
markup. OfficeAgent.NET handles that document-specific work. The model works
with structured document data and JSON-serialisable operations such as "replace
this clause as a tracked change" or "add a row to this table."

The same engine is available in three forms:

• an MCP server for agents that support the Model Context Protocol;
• tools for Microsoft Agent Framework and Microsoft.Extensions.AI;
• a .NET API for applications that want to control the workflow directly.

It supports Word .docx files and PowerPoint .pptx decks; one client serves
both, routing each document to the module that handles it. Excel is not
implemented. See Scope and limitations before choosing
it for a workflow that depends on Office's layout or calculation engine.

What you can build

Choose a starting point

Try a Word edit

This small workflow demonstrates that OfficeAgent can change an existing OOXML file
without flattening its structure. It uses tracked changes because the result is easy to
verify in Word; review is one part of the broader document operation set.

Install the server. The published package command is:

dotnet tool install --global OfficeAgent.Mcp

Make a folder for the agent to work in and download the
sample contract into it — a fictional services
agreement with a clause to change, a table, an open comment, and a pending redline:

mkdir -p ~/officeagent-documents
curl -Lo ~/officeagent-documents/services-agreement.docx \
  https://raw.githubusercontent.com/ilia-sokolov/OfficeAgent.NET/main/samples/documents/services-agreement.docx

PowerShell:

$officeAgentDocuments = Join-Path $env:USERPROFILE "officeagent-documents"
New-Item -ItemType Directory -Force $officeAgentDocuments | Out-Null
Invoke-WebRequest `
  https://raw.githubusercontent.com/ilia-sokolov/OfficeAgent.NET/main/samples/documents/services-agreement.docx `
  -OutFile (Join-Path $officeAgentDocuments "services-agreement.docx")

Any .docx of your own works too — the sample just gives you something with a comment and a
pending revision already in it.

Register the server with Claude Code, pointed at that folder and nothing else:

claude mcp add \
  --env OfficeAgent__FileSystemConnections__0__ConnectionId=documents \
  --env OfficeAgent__FileSystemConnections__0__RootPath=$HOME/officeagent-documents \
  --transport stdio \
  officeagent -- officeagent-mcp --stdio

PowerShell:

claude mcp add `
  --env OfficeAgent__FileSystemConnections__0__ConnectionId=documents `
  --env "OfficeAgent__FileSystemConnections__0__RootPath=$officeAgentDocuments" `
  --transport stdio `
  officeagent -- officeagent-mcp --stdio

For this review-specific workflow, you can optionally install the
word-document-review skill before starting the client.

Then ask:

In services-agreement.docx, change the payment terms from thirty days to forty-five days.

Open the file in Word. Clause 3 now reads forty-five days as a tracked change you can
accept or reject, and everything else — the table, the comment, the redline that was
already there — is exactly as it was. This demonstrates a key engine property: apply the
requested operation while preserving unrelated package content.

What else the sample is good for — reviewing comments,
accepting revisions, editing the table.

Next, try creating a Word document,
generating a PowerPoint deck, or using the
direct .NET workflow.

If it does not work

Configure broader workflows

The quick start above is deliberately the smallest thing that works. Four settings extend it:

Past a couple of settings, use a file instead — the same OfficeAgent section, where a list
is a list:

{
  "OfficeAgent": {
    "AllowCreation": true,
    "FileSystemConnections": [
      {
        "ConnectionId": "documents",
        "RootPath": "C:\\officeagent-documents",
        "AllowedExtensions": [ ".docx", ".pptx" ],
        "DefaultChangeMode": "Direct"
      }
    ]
  }
}

The same configuration is available as
samples/config/word-and-powerpoint.json. Change
RootPath before using it.

claude mcp add --transport stdio officeagent -- officeagent-mcp --stdio --config ./officeagent.json

Environment variables still override the file. Windows, PowerShell, other MCP clients,
HTTP hosting and SharePoint are in
Deployment and client setup; every setting is listed in
MCP server.

Optional guidance for Word review

skills/word-document-review teaches the review
loop: read comments and pending revisions before editing, keep reviewable Word edits as
redlines, use document ids for multi-step work, and recover from stable error codes. The
installation guide gives complete Bash and PowerShell steps
for Claude Code and Codex, including installation from a fresh machine and verification.
The skill is only needed when the task requires that review discipline; document creation,
ordinary direct edits, and PowerPoint workflows use the server without it.

What reaches the model

The inspect and find tools return document text and structure to the model — that is how it
locates an edit. Filesystem and SharePoint operations keep the package behind an opaque id.
Inline tools carry the whole file as base64 on every call. Session import/export also carries
the package as base64 if the agent performs those calls; a host integration can instead move
the bytes outside model context. Connect storage and model providers appropriate for the data.

The server ships no authentication layer for HTTP hosting; put it behind your own. A
filesystem root is a trust boundary: its ACLs must stop untrusted principals creating,
renaming or replacing entries while the server runs.

.NET quick start

Install the core package and Word module:

dotnet add package OfficeAgent.Core
dotnet add package OfficeAgent.Word

After registering services and a document provider, the edit loop looks like
this:

var client = services.GetRequiredService<OfficeAgentClient>();
var doc = await client.RegisterAsync("workspace", "/srv/workspace/contract.docx");

var inspect = await client.InspectAsync("workspace", doc.ItemId);
var hit = (await client.FindAsync(
    "workspace", doc.ItemId, new FindQuery("Acme Corp"))).First();

var plan = new DocumentPlan
{
    Snapshot = inspect.Snapshot,
    Operations = new PlanOperation[]
    {
        new ChangeTextOp
        {
            Target = hit.Anchor,
            With = "Globex Inc.",
            Mode = ChangeMode.Tracked
        }
    }
};

var preview = await client.PreviewAsync("workspace", doc.ItemId, plan);
if (preview.IsValid)
    await client.CommitAsync("workspace", doc.ItemId, plan);

The complete example, including service registration and reading the saved
file, is in Getting started.
The minimal sample replaces the first Acme Corp with Globex Inc.. To run it,
copy a Word document containing Acme Corp to contract.docx in the cloned
repository root, then run:

dotnet run --project samples/QuickEdit -- ./contract.docx ./contract-edited.docx

The repository also contains a
direct IChatClient Word-editing sample and an
interactive
Agent Framework sample, plus a complete
contract-review agent that separates model judgement from
validated document writes.

How it works

Every edit follows the same four steps:

1. Inspect returns a structured map of the document: its outline,
   paragraphs, styles, content controls, tables, images, and revisions.
2. Find searches text and returns a content-verified anchor for each match.
3. Preview validates a plan against the current document and reports the
   proposed changes without writing.
4. Apply commits the complete plan and saves it through the configured
   provider.

A plan (DocumentPlan) is a typed, JSON-serialisable list of operations. An
anchor records both a location and the content expected there. If the content
or optional document snapshot has changed, validation fails instead of silently
targeting a different location. Applying a plan is all-or-nothing.

The Word module supports changes to text, paragraphs, tables, images, styles,
content controls, comment threads, footnotes and endnotes, page geometry and
breaks, document properties, and tracked revisions. Operations with a Word revision
representation record a redline when the connection asks for one - an inserted clause,
a deleted row and a restyled heading all come back as revisions a reviewer
accepts or rejects, not only a replaced phrase. Image resizing is applied directly because
WordprocessingML has no revision representation for drawing dimensions. The
PowerPoint module implements a broad, explicitly documented set of deck
operations: text, bullets, run and paragraph formatting, template
slots, style copying, tables, images, text boxes, embedded video and audio,
speaker notes, resolvable comments, footers and slide numbers, sections,
transitions and animations, and the slide lifecycle - adding, removing,
reordering and duplicating. Several slide inserts in one plan author a deck end
to end, so a single call turns nothing into a finished presentation. Any verb it
does not support is named rather than silently skipped. The full operation
schema is documented in Document plans, and the deck
specifics in PowerPoint support.

Documents are accessed through configured providers. After registration,
editing calls use a (connectionId, documentId) pair instead of a storage path
or credentials. The filesystem provider restricts registrations to its root;
the SharePoint provider uses the permissions of its configured identity.
CreateAsync starts a new document inside a connection: the requested .docx
or .pptx extension selects a registered blank-document factory. The engine
applies an optional initial plan in memory, and then asks
the provider to create and register it without overwriting an existing name.

Documentation

Contributing

Bug reports, documentation fixes, new document operations, provider
integrations, and focused test cases are useful contributions. If you found a problem,
open an issue with the
document feature involved, the operation you attempted, and the error or
unexpected result. Do not attach confidential documents; a small sanitised
reproduction is enough.

To work on the code, install the .NET 8 SDK, fork the repository, and run:

dotnet build OfficeAgent.NET.sln
dotnet test OfficeAgent.NET.sln

Before starting a larger change, especially one that changes public types or
the JSON wire format, open an issue
so the design can be discussed. See
CONTRIBUTING.md
for code style, tests, and pull-request expectations.

Scope and limitations

OfficeAgent.NET edits Word .docx files and PowerPoint .pptx decks; it does
not automate the Office desktop applications. An Excel module can be added
through IFormatModule, but it does not ship today.

The deck module refuses the verbs a presentation has no vocabulary for -
setProperty, revision, pageSetup, insertBreak and note - per
operation, rather than applying part of a plan, and refuses an explicit tracked
mode on any verb that carries one. PresentationML has no redline model, so tracked changes are Word-only, and
a slide has no header (that is a notes and handout concept). Animations cover
the effects expressible as a filtered p:animEffect; fly-in, zoom and motion
paths are refused rather than approximated. See
PowerPoint support for what a deck does and does not
accept.

The engine does not render pages or calculate Word fields. Operations that
depend on pagination, table-of-contents rendering, field recalculation, or
page-fit checks are outside its scope. Preview reports structural changes, not
a visual rendering of the final document. Test the workflow on representative
documents and keep human review in the loop for consequential edits.

Two more limits worth knowing before you build on it:

• Token savings depend on how you connect. Addressing a document by id keeps
  the package out of the conversation, and inspection can be narrowed with
  fidelity and paging - that is where the saving comes from. The inline
  *_content tools are the deliberate exception: they carry the whole file as
  base64 in both directions, which costs tokens in proportion to file size. They
  suit a single self-contained call, not a sequence of edits - a model asked to
  pass a document of a few kilobytes back for a second edit reproduces it
  imperfectly and the follow-up fails. Use a connection, or a session connection,
  when more than one edit is coming.
• Review guidance is optional. For review tasks, the server alone does not make an
  agent read open comments before editing or choose a redline. The
  word-document-review skill teaches that workflow;
  without it, review behaviour depends on the model and the prompt.

Commercial support

OfficeAgent.NET is MIT-licensed and can be self-hosted. Managed hosting and
commercial support are available from dotaction:
contact dotaction.

License

MIT. See LICENSE.