MATLOOK AgentUser guide
中文Open web app

Projects & files

Attachments, folders & outputs

Know where inputs are read, outputs are saved and which files are available remotely.

View Markdown ↗

Distinguish three locations#

Attachments are inputs to the conversation. The working folder provides a context for file tools. The project asset directory stores project materials. These can be different paths.

Add and use references#

  1. Preserve clearly named originals with revision identifiers, such as “demo-materials-v1.csv”.
  2. Attach files in the composer or enable a suitable local-folder source.
  3. Ask the model to confirm which files it read and list parsing failures separately.
  4. Specify a destination and new output filename so the original is preserved.
  5. Open the actual output using its result link and inspect its contents.

Does readability guarantee understanding?#

No. Scanned drawings, text in images and complex tables need review; image understanding also depends on the selected model. Cross-check specifications, units, dimensions and material IDs against the original. If parsing fails, provide clearer page images or a document with selectable text.

What happens remotely?#

A computer's path is meaningful on that computer. A filename or link in mirrored history does not mean the original is available on the phone or browser. Attachments and full tool outputs may require the execution computer. See remote access.

Acceptance checks#

Verify that the file exists, opens and has the correct name and revision. Check units and row counts in tables, sources in reports and clarity in images. A completion message does not replace inspection of the deliverable.

Built-in document-tool failures (unreleased repair preview)#

The repair preview preserves the original error and exit status, with tool-reported interpreter type, Python version, allowed dependency versions and failure stage. Argument validation is separate from Python exceptions. Metadata excludes script contents, file paths and credentials, and does not change permissions or execution results. External or user-configured Python is not automatically installed or repaired; unknown runtimes remain unknown. Historical errors still require their original evidence and are not considered fixed based on their class alone.

Local attachment preparation (rc.75 beta)#

This release provides Mac Apple Silicon and Windows x64 desktop packages. Files used through the web or phone are still processed on the connected execution computer. Windows 11, the final installer upgrade/uninstall lifecycle and a real-material extended trial remain unverified. No fixed traffic-saving percentage is promised.

Complete originals are saved in the session attachment folder on the execution computer. Large text previews may be shortened while saved originals remain complete. Resized images have a separate analysis copy; clicking an attachment opens the original.

Word, Excel, PowerPoint and PDF files reuse offline parsers and receive section indexes. PDFs use embedded text first and local Chinese/English OCR for scanned pages. Spreadsheet evidence retains cell addresses, formulas and cached values without running macros or recalculating formulas. Attachment cards show indexed, partial, failed or local-tool status. Before sending, the app shows saving, queued and parsing stages; scanned PDFs show page progress. Long parsing jobs can be cancelled without deleting originals. After failure or cancellation, retry the saved original, send saved originals, or return to editing. Pending text and attachments are retained in the draft. Files that could not be saved are reported separately.

The conversation can search and read selected sections with persistent delivery coverage. Small files (up to 4 sections) can be read in one call without a preliminary overview. Larger files remain bounded per read, and whole-document analysis still requires every section. Extracted spreadsheet coordinates, stored formulas and cached values can be used directly as text evidence; inspect originals for missing or ambiguous evidence, layout and charts. Search snippets do not count as full reads. Coverage records text returned by the tool, not model comprehension, visual inspection or formula verification. Whole-document analysis should cover every section and disclose gaps. Cross-check specifications, colors, textures, drawings and complex tables against originals. Older attachments remain readable at their original paths.

Files and parsing remain on the execution computer. When using an online model, selected text, visual inputs and tool results still travel to the model. Native PDF/image behavior remains compatible with each model backend. File references do not guarantee that later history avoids retransmission.

Based on MATLOOK Agent functionality, bundled help and release notes. Available controls depend on the installed version, execution target and account permissions. Versions & editorial notes ↗
Search the full guide · Esc to close