ESC
Type to search across all documentation
Tools

Tools

Tools are what the AI uses to interact with your system. When the AI reads a file, runs a command, or searches the web, it's using a tool. You'll see tool activity as inline status indicators during the conversation.

File Operations

read_file

Read the contents of any file. In Normal mode, reading files outside your working directory requires approval.

write_file

Create a new file or overwrite an existing one. Parent directories are created automatically. Requires approval in Normal mode. Blocked in Plan mode.

edit_file

Make a targeted edit to an existing file by finding and replacing a specific string. The search string must appear exactly once in the file. Requires approval in Normal mode. Blocked in Plan mode.

glob

Find files matching a pattern (e.g., **/*.ts, src/**/*.test.js). Ignores node_modules, dist, and .git by default. Results are limited to 100 matches.

grep

Search file contents using regular expressions or plain text. Skips binary files and files over 1MB. Results capped at 100 matches. Ignores node_modules, dist, .git, and vendor by default.

list_dir

List directory contents with file sizes. Can optionally show hidden files and recurse up to 3 levels deep. Automatically skips node_modules, vendor, .git, dist, and build.

Execution

bash

Execute shell commands. Requires approval in Normal mode. Blocked in Plan mode. Output is capped at 50,000 characters.

python

Execute Python 3 code. Requires python3 to be installed on your system. No approval required in any mode.

applescript

Execute AppleScript for macOS automation. Only available on macOS. No approval required in any mode.

Web

Search the web via DuckDuckGo. No approval required in any mode. Limited to 5 calls per response to prevent search loops — the AI is instructed to synthesize results after 2–3 searches rather than continuing indefinitely.

http_request

Fetch content from any URL. Supports GET, POST, PUT, and other HTTP methods. HTML responses are automatically cleaned to plain text. No approval required in any mode.

Task Tracking

todo_update

Update the visual task list displayed in the interface. The AI uses this to track progress on multi-step tasks, showing items as pending, in progress, or completed.

Mode Restrictions

Tool Normal Auto Plan
read_file (in working dir) Allowed Allowed Allowed
read_file (outside working dir) Approval Allowed Allowed
write_file Approval Allowed Blocked
edit_file Approval Allowed Blocked
bash Approval Allowed Blocked
python Allowed Allowed Allowed
applescript Allowed Allowed Allowed
glob, grep, list_dir Allowed Allowed Allowed
web_search, http_request Allowed Allowed Allowed
todo_update Allowed Allowed Allowed

Safety Limits

Blocked Commands

The following shell commands are always blocked, regardless of mode:

  • rm -rf / and rm -rf ~
  • mkfs (filesystem formatting)
  • dd if= (disk writes)
  • Fork bombs
  • > /dev/sda (raw disk writes)
  • chmod -R 777 / (recursive global permissions)
  • chown -R (recursive ownership changes)

Per-Session Tool Limits

Tool Max Calls
read_file 500
glob, grep, list_dir 100 each
edit_file, todo_update 100 each
bash 50
write_file 50
python 30
http_request 10
web_search 5

The AI can make up to 50 tool iterations per response before stopping.

Image Generation

LogiCoal can generate images and SVG vector graphics from text descriptions using FLUX.

When you ask for an image, your prompt is automatically enhanced by the reasoning model before being sent to FLUX. The enhancer:

  • Rewrites vague requests into detailed visual descriptions optimized for FLUX
  • Resolves contextual references (e.g., "generate a logo for it") using recent conversation history
  • Removes text/label instructions since FLUX cannot render readable text
  • Adds art style keywords (flat design, photorealistic, vector style, etc.)
  • Skips enhancement for prompts over 300 characters (assumed to be already detailed)

For SVG generation, the enhanced prompt generates a FLUX image which is then vectorized into 4 SVG candidates with different presets.

Content Moderation

The server-side task classifier detects requests for harmful content — malware, game exploits, anti-cheat bypasses, phishing tools, keyloggers, and similar. These requests are declined by the reasoning model with a brief explanation and suggested alternatives, in whatever language the user is writing in.

Legitimate uses of the same keywords are not blocked. For example, "fix SQL injection in my code" routes normally, while "write a SQL injection tool" is declined.

Auto-Open (macOS)

When the AI generates or saves image files during a session, LogiCoal automatically opens them in Finder so you can preview the result immediately. This applies to PNGs, SVGs, and other image formats.