26.10.0

Initial release of GroupDocs.Editor.Mcp

What changed

• New MCP server for GroupDocs.Editor for .NET 26.6.1 on GroupDocs.Mcp.Core 26.9.0, cloned in
  shape from GroupDocs.Conversion.Mcp 26.9.0 (workflows, install/ generator, Docker, build and release scripts).
• Seven tools, plus Core's get_license_status:
  • get_document_info — family, format, pages/worksheets/slides, size, encryption, save targets (JSON).
  • extract_html — a document, or one worksheet, slide or PDF page range, to editable HTML in storage;
    self-contained by default or with a <name>_resources folder; plain-text preview.
  • save_from_html — HTML back to DOCX, XLSX, PPTX, PDF, TXT, Markdown, EPUB and more; splices one
    edited worksheet or slide back into sourceFile; optional password and output name.
  • replace_text — find/replace on decoded text nodes inside <body> (literal, case-insensitive,
    whole word, regex), one worksheet or slide at a time, optional output format; zero matches writes nothing.
  • get_supported_formats — load/save catalog per family, or save targets for one file.
  • get_form_fields / fill_form_fields — legacy Word form fields; unknown names and unfit values reported.
• Tools return CallToolResult: every failure sets isError and starts with the tool's prefix
  (HTML extraction failed for '<file>': …). Temp paths are scrubbed from engine messages.
• Licensing: EditorLicenseManager implements licence file, metered key and consumption reading, and
  declares the engine for get_license_status.
• Package: nuget-first, 174 MB nupkg (measured), McpServer package type, native PDBs stripped.

Decisions

• Native dependencies: none beyond fonts. GroupDocs.Editor 26.6.1 uses Aspose.Drawing and declares
  SkiaSharp + SkiaSharp.NativeAssets.Linux.NoDependencies itself, so there is no
  System.Drawing.EnableUnixSupport, no libgdiplus, and no manual SkiaSharp pin. The Docker image
  installs libfontconfig1, fontconfig (for fc-cache) and ttf-mscorefonts-installer for PDF/XPS output.
• Security pin: System.Security.Cryptography.Pkcs 10.0.13 overrides the engine's transitive 8.0.0
  (GHSA-846f-f764-76mx, GHSA-2fmv-rv2c-59x7).
• Docker on Linux: README and Docker Hub README document --user "$(id -u):$(id -g)" from day one. A write the
  container user is denied (the same cause) fails with that fix in the message, not only the raw
  UnauthorizedAccessException — the lesson of GroupDocs.Search.Mcp 26.10.1.

Why

AI agents are being asked to change Office files "without wrecking the formatting"; GroupDocs had no
MCP server for its editing engine. This one targets that intent directly.

Fixed before the first release (code review 2026-10-29)

Every item below was reproduced against the local build first.

• Evaluation mode never edits the wrong part. Unlicensed, the engine returns slide 2 for slide 3 (and
  worksheet 2 for worksheet 3) and reports success; replace_text would then have spliced that edit back.
  extract_html and replace_text now refuse worksheet/slide 3+ without a licence and say why.
• Hidden worksheets. Worksheet numbers count every sheet, as the save side does; for Office Open XML
  workbooks hidden sheets are mapped around and a hidden worksheet is refused rather than silently swapped
  for worksheet 1.
• save_from_html splicing. A worksheet/slide number beyond the source is refused (the engine replaced
  the last part); a workbook sourceFile needs worksheet; a presentation sourceFile without slide is
  reported as a new one-slide deck, not "a copy of" the source; a sourceFile, worksheet or slide that
  does not fit the target is refused.
• extract_html with embedResources=false reserves the .html name in the folder storage really writes
  to, so a second extraction gets name (1).html + name (1)_resources/ instead of overwriting the first
  extraction's resources (wrong when GROUPDOCS_MCP_OUTPUT_PATH was set).
• replace_text keeps a protected input's password in any output that can carry one (PDF included) and
  says plainly when the target cannot; refuses an input it cannot save back (HTML, XML) before editing; and
  explains the evaluation PDF page cap up front instead of failing with "the file cannot be opened".
• Regex find patterns get a 15 s budget for the whole document, not 5 s per text node.
• Engine messages name the file they are about (sourceFile vs the HTML), and tool descriptions no longer
  promise the failure prefix for a missing file (that case returns the list of available files).

Verification

• Unit tests: 93/93 (dotnet test src/GroupDocs.Editor.Mcp.sln).
• Integration suite (GroupDocs.Editor.Mcp.Tests, 80 tests) in evaluation mode: 80/80 against this build via
  MCP_SERVER_DLL on Windows, and 80/80 against the locally built image via MCP_SERVER_IMAGE (Linux container).
• Installer verifier (verify-groupdocs-mcp.ps1, handshake + auto) against the local image; a sequenced
  dnx session against the packed .nupkg.
• build.ps1: server.json lockstep holds; install/generate-install-links.ps1 -Check clean.

Migration / impact

None — first release.

Full Changelog: https://github.com/groupdocs-editor/GroupDocs.Editor.Mcp/commits/26.10.0