Documents, Images & Web
TheGitAI can bring specifications, spreadsheets, screenshots, diagrams, and current public information into the same session as your source code.
Documents
The structured document reader supports:
| Format | Read | Edit |
|---|---|---|
| Yes, with bounded page ranges | No | |
| XLSX | Yes, including workbook sheets | No |
| DOCX | Yes, with bounded paragraph ranges | Yes, visible text replacement |
Normal source, text, Markdown, JSON, YAML, and configuration files use the regular file tools rather than the document reader.
Editing Word documents
DOCX editing replaces visible text while preserving the document package. It can match text in paragraphs and tables, apply several replacements in one operation, and write either to the original path or a separate output path.
Before writing, the agent can validate the requested matches without changing the file. The terminal shows a preview for approval, and the result reports each replacement so partial matches can be corrected precisely. DOCX writes are also captured by the assistant edit journal and checkpoint system.
requirements.pdf, compare
it with the current API, and update implementation-plan.docx where the old
endpoint names appear.”Images and screenshots
Attach an image by pasting it from the clipboard or by typing its path in the
prompt. Paths with spaces may be quoted. The terminal replaces detected image
paths with stable markers such as [Image #1] so you can refer to each image
unambiguously.
Supported formats are PNG, JPEG, GIF, and WebP. A message can contain up to two images, with a 10 MB limit per image.
Vision-capable models receive the image with your prompt. TheGitAI can use images to:
- Inspect a UI screenshot or error dialog.
- Read visible labels, controls, and layout.
- Compare an implementation with a reference image.
- Understand diagrams and other visual task context.
The exact clipboard shortcut is shown in the terminal footer for your platform. On systems where clipboard image access needs a helper, the terminal explains which local utility to install.
Web search
The agent can search the public web when a task needs current or external information. Search results include titles, URLs, and snippets. Independent searches may run in parallel with other read-only research.
Web fetch
The fetch tool retrieves one public HTTP or HTTPS URL and converts the useful page content to Markdown. It reports coverage, visible inventories or listings, and truncation details rather than silently pretending a partial extraction was complete. Large pages include continuation information so the agent can read the next bounded section.
For safety, the managed fetch tool rejects localhost, private IP ranges, and internal network destinations. When the task is explicitly about a local dev server, the agent uses an approved local shell command instead.
Source quality still matters
Web results and extracted document text are evidence, not guaranteed truth. Ask the agent to compare multiple sources, inspect the primary documentation, or state what remains uncertain when accuracy is important.