Tools & Actions

TatsuCode can take practical actions in your workspace. You describe the goal; it chooses the right action path.

This page is intentionally product-focused and avoids low-level internal implementation details.

What TatsuCode Can Do

File Operations

Typical requests:

Explain @src/auth/service.ts
Create a utility for date parsing
Fix error handling in this function
Find all TODO comments in the project

You can also drag files/folders into input for quick context loading.

Search and Find & Replace

Typical requests:

Find every place we call the payment API
Find invoice.pdf anywhere on my drives
Rename OldService to BillingService across the project
Add a copyright header to every source file

Search finds text inside files, or files and folders by name, size, or date, in the project, any folder, or every drive, usually in a fraction of a second. Long searches keep running in the background with a live card and a Stop button. Task Agents search with the same engine. See Search.

Find & Replace previews every edit before anything is written, leaves out what shouldn't change, and undoes a whole batch in one step. See Find & Replace.

Task Checklist

Typical requests:

Keep a checklist while you work on this
Plan the upgrade as a checklist first, then start

For work with several steps, the agent keeps its plan in one live checklist card: done, in progress, and up next, with a progress bar. Claude Code fills it from its own task tools; every other model uses TatsuCode's Update Plan tool, and a plan carries over when you switch models. See Task Checklist.

Office Document Work

Typical requests:

Read this DOCX and summarize action items
Update this XLSX range with corrected numbers
Replace placeholder text in this PPTX
Create a new DOCX from this markdown outline

This is useful for report updates, spreadsheet cleanup, and presentation prep without leaving TatsuCode.

Code Analysis

Typical requests:

Find where this class is defined
Show all references to this method
Check this project for lint/build issues

Use this when diagnosing breakages, refactoring safely, or validating changes before merge.

Terminal Work

Typical requests:

Run tests
Build the project
Install missing dependencies

TatsuCode requests approval for sensitive operations based on your permission settings.

When an agent reaches for a shell command to search through files (grep -r, findstr /s, dir /s, and the like) or to rewrite text across many files (sed -i, a PowerShell -replace loop), TatsuCode points it to Search or Find & Replace instead, with a ready-made equivalent: they're faster, and a preview shows what will change. Claude Code gets a short reminder rather than a redirect. When the shell really is the right tool, the agent can still run the command.

Web and Browser Work

Typical requests:

Search for migration guidance for this framework
Open my local app and click through login flow
Check browser console for errors
Take a screenshot after the page settles

This is especially useful for UI debugging and frontend QA loops.

YouTube Analysis Work

Typical requests:

Get transcript for this YouTube video
Analyze this YouTube tutorial and list implementation steps
Do deep analysis of this video and summarize tradeoffs

Use transcript mode for speed, and video analysis modes when visual/audio details matter.

Undo and File Recovery

Typical requests:

Undo the last edit to this file
Restore this file to an earlier state
Show what changed in this session

Use recovery tools whenever a change goes in the wrong direction.

Desktop Automation (Computer Use)

Typical requests:

Click the "Save" button in the settings window
Fill out the registration form in this desktop app
Take a screenshot, find the error dialog, and close it

Computer Use lets TatsuCode interact with any desktop application — clicking buttons, typing text, dragging items, scrolling, and managing windows. Load the computer-use skill to enable it.

Screen Recording

Typical requests:

Record the browser while I click through the login flow
Start recording the app window and watch for visual bugs
Record, export, and analyze what happened during that test

Screen recording captures video clips from windows or the full screen. The AI can then watch the recording to give feedback or verify behavior.

Image Generation, Editing, and Annotations

Typical requests:

Generate a pixel-art potion bottle sprite sheet
Take @img-02 and replace the sky with a dramatic sunset
Resize this screenshot to 1024x1024 and trim the transparent edges
Crop @img-04 to just the error dialog

Configure the image generation provider, model, and defaults via /image-generation. Browse everything that's been generated, attached, or screenshotted via /image-gallery. Mark up images directly in chat (circle, point, note) and the agent receives the annotated version.

See Images for the full workflow.

Elapsed-Time Tracking

Long agent runs now show how long the task took to finish, and total work time is included in /usage. Useful for getting a feel for cost and for comparing model speed on the same task.

Plugin Integrations

Typical requests:

Connect to Unity and create a new empty scene
Add a directional light and a camera to the scene
Connect to Maya and rig this character
Close Blender when you're done

Plugins connect TatsuCode to live editors: Unity, Unreal Engine, Godot, Blender, and Autodesk Maya. Ask the agent to connect and it installs the bridge and connects by itself; then it builds scenes, materials, Blueprints, rigs, and renders directly from chat. Use /plugins to browse and connect. See Game Engine Plugins.

Typed Decisions (TypeSafe AI)

Typical requests:

Triage these failing tests into flaky, regression, or environment
Rank these candidates by fit for the ticket and keep the top five
Is this command destructive? Check before running it
Wire an NPC in my Godot scene to pick its next action with TypeSafe

With the typesafe-ai skill loaded, the agent can ask TypeSafe AI's decision model typed questions about a piece of state (choose one of many, place on a scale, yes or no) and get constrained answers with confidence in well under a second. It uses the tool inside a task and writes the same calls into your code when you ask for a feature. Bring your own key via /connect → Decision Models. See Decision Models.

Cross-Talk Between Windows

Typical requests:

List the other TatsuCode windows
Send the unreal-sandbox window a note that the material names changed
Ask the other window to run the tests and message me the failures

Every window has an address and a Peers tool with list, send, status, and whoami. A message arrives in the other window as a distinct Message from block and its model acts within that window's own permissions. Works across providers: Gemini to Claude Code, local model to OpenRouter, and back. Loopback only, per-window tokens. See Cross-Talk.

Agent Skills in Tool Workflows

Use skills when tasks need a specific method or quality bar.

/skills

Examples:

• Load deep-think for difficult debugging or design tradeoffs
• Load skill-editor to create or improve reusable team skills

See Agent Skills.

Permission and Safety Model

Use:

/permissions

Manage permission policy and grant individual command-line tools — permanently or for the current session, including unknown tools the agent has never run before.

Set each category to:

• Ask
• Allow
• Deny

Recommended: keep sensitive categories on Ask unless you fully trust the task context. For power users, see YOLO mode in Permissions and YOLO.

Best Practices

• Describe outcomes, not implementation steps
• Keep requests focused and scoped
• Review diffs before approving edits
• Compact long sessions when context gets large
• Use skills for repeatable workflows

Next Steps

• AI Capabilities — broader feature overview
• Agent Skills — specialized workflow system
• Commands — slash command reference
• Troubleshooting — issue recovery guide