Read a document and convert it to Markdown format (simplified API).

Supports PDF, Word (DOCX), PowerPoint (PPTX), HTML, email (EML), EPUB, Jupyter Notebooks (IPYNB), Excel (XLSX/CSV), ODF formats, and 200+ text formats.

The source parameter is auto-detected as:

• File path (if file exists in read allowlist)

• Data URI (data:image/...)

• Base64-encoded content (auto-detected)

• Plain text content (HTML, Markdown, etc.)

Optionally extract a specific section by providing the section parameter with a heading name (case-insensitive match).

Image inclusion and markdown flavor are configured at server startup and cannot be changed per-call.

Returns the Markdown as a text content block, followed by an image content block for each image found (when include_images=true), so vision models can see the images alongside the text. When include_images=false, only the text, with image alt text in place of the images.

Search a corpus of documents and return ranked snippets.

Instead of returning whole files, this tool returns the most relevant passages (with hit spans wrapped in << >> markers), so an agent can locate information across many documents cheaply.

Modes:

• keyword: BM25 relevance ranking (requires the optional rank-bm25 dependency). Best for "find the most relevant passages" queries.

• grep: literal/regex line matching. Best for "find every occurrence of X". Stateless, no extra dependencies.

Supports all formats the read tool supports (PDF, DOCX, HTML, etc.); files are converted to text before searching.

Returns a dictionary with:

• results: list of {snippet, score, document_path, section_heading, chunk_id}

• mode: the search mode used

• total: number of results returned

Compare two documents and return their differences.

Each input is auto-detected (file path, data URI, base64, or inline content), so documents in any supported format can be compared — even across formats (e.g. a DOCX against its PDF export).

Returns a dictionary with:

• diff: the rendered diff (unified text or JSON string)

• has_changes: whether any differences were found

Return the heading structure (table of contents) of a document.

Use this to navigate a large document before extracting specific sections: the returned indices line up with edit_document's '#N' target notation.

Returns a dictionary with:

• sections: list of {index, level, heading}

• total: number of headings returned

List files the server is allowed to read.

Use this to discover what is available before reading or editing. Results are confined to the read allowlist (the workspace folder plus any additional read-only folders). Paths returned here can be passed directly to the other tools.

Returns a dictionary with:

• files: list of {path, size_bytes} (absolute paths, sorted)

• total: number of files returned

• truncated: true if more files exist than were returned (capped)

• read_dirs: the read folders that were searched