dhalsim

Browser automation gadgets for llmist agents using Camoufox anti-detect browser.

Using with llmist CLI

Use dhalsim gadgets directly from the command line for quick tasks and testing.

Quick Start

# Use the BrowseWeb subagent
bunx @llmist/cli agent "go to apple.com and find iPhone 16 Pro price" -g dhalsim:subagent

# Use all gadgets (for custom agent workflows)
bunx @llmist/cli agent "navigate to example.com" -g dhalsim

# Use readonly preset
bunx @llmist/cli agent "take a screenshot of google.com" -g dhalsim:readonly

# Use latest dev from GitHub (with BrowseWeb subagent)
bunx @llmist/cli agent "search google for llmist" -g git+https://github.com/zbigniewsobiecki/dhalsim.git#dev:subagent

Configuration

Configure BrowseWeb subagent in ~/.llmist/cli.toml:

[subagents.BrowseWeb]
model = "sonnet"              # LLM model for the subagent (default: sonnet)
maxIterations = 20            # Max agent loop iterations (default: 15)
headless = true               # Run browser in headless mode (default: true)
timeoutMs = 600000            # Overall timeout in ms (default: 300000 = 5 min, 0 = disabled)
disableCache = false          # Disable browser cache for lower memory usage (default: false)
navigationTimeoutMs = 60000   # Navigation timeout in ms (default: 60000)

Per-profile configuration

[develop.subagents.BrowseWeb]
headless = false           # Show browser during development

[research.subagents.BrowseWeb]
maxIterations = 30         # More iterations for deep research

Using "inherit" for model

[subagents.BrowseWeb]
model = "inherit"          # Use parent agent's model

Production Deployment

For production environments, consider the following settings:

[production.subagents.BrowseWeb]
headless = true               # Always run headless in production
disableCache = true           # Reduces memory per browser instance (~100-200MB savings)
navigationTimeoutMs = 90000   # 90s for slow networks or heavy sites
maxIterations = 20            # More iterations for reliability

Resource requirements:

• Each browser instance uses ~300-500MB RAM
• Running multiple concurrent BrowseWeb calls multiplies memory usage
• Use disableCache = true for environments with <1GB RAM per browser

Custom Commands in cli.toml

[my-research-command]
gadgets = [
  "dhalsim:subagent",                                                  # from npm
  "git+https://github.com/zbigniewsobiecki/dhalsim.git#dev:subagent",  # from git
]

Using in Projects

Install dhalsim as a dependency and use gadgets programmatically.

Installation

npm install dhalsim
# or
bun add dhalsim

Using Dhalsim Subagent

import { LLMist } from 'llmist';
import { Dhalsim } from 'dhalsim';

const result = await LLMist.createAgent()
  .withModel('sonnet')
  .withGadgets(new Dhalsim())
  .askAndCollect('Go to google.com and search for "playwright"');

// With custom timeout (10 minutes)
const dhalsim = new Dhalsim({ timeoutMs: 600000 });

// Disable timeout for debugging
const debugDhalsim = new Dhalsim({ timeoutMs: 0 });

Using Individual Gadgets

import { LLMist } from 'llmist';
import { createGadgetsByPreset } from 'dhalsim';

const gadgets = createGadgetsByPreset('all');

const agent = LLMist.createAgent()
  .withGadgets(...gadgets)
  .ask('Navigate to example.com and take a screenshot');

for await (const event of agent.run()) {
  // handle events
}

Presets

• all (default) - All gadgets
• readonly - Navigate, Screenshot, GetFullPageContent, ListPages
• minimal - Navigate, Screenshot, GetFullPageContent
• subagent - BrowseWeb subagent only

Available Gadgets

| Category | Gadgets | |----------|---------| | Page | NewPage, ClosePage, ListPages | | Navigation | Navigate, GoBack, GoForward, Reload | | Content | GetFullPageContent, Screenshot | | Interaction | Click, Type, Fill, FillForm, PressKey, Select, Check, Hover, Scroll, DismissOverlays | | Script | ExecuteScript | | Wait | WaitForElement, Wait | | User Input | RequestUserAssistance |

Subagents

BrowseWeb

Autonomous browser agent that can navigate, interact, and extract information from websites. Runs its own agent loop internally, making it suitable for complex multi-step web tasks.