Skip to main content

Documentation Index

Fetch the complete documentation index at: https://stagehand-shrey-check-v3-metrics-docs.mintlify.app/llms.txt

Use this file to discover all available pages before exploring further.

Stagehand provides built-in action caching to reduce LLM inference calls and improve performance. Simply specify a cacheDir when initializing Stagehand, and actions are automatically cached and reused across runs.

​
How Caching Works

When you specify a cacheDir:
  1. First run: Actions use LLM inference and results are cached to a local file
  2. Subsequent runs: Cached actions are reused automatically (no LLM calls)
  3. Cost savings: Eliminate redundant inference calls for repeated actions
  4. Performance: Faster execution by skipping LLM inference

​
Caching with act()

Cache actions from act() by specifying a cache directory in your Stagehand constructor. 
import { Stagehand } from "@browserbasehq/stagehand";

const stagehand = new Stagehand({
  env: "BROWSERBASE",
  cacheDir: "act-cache", // Specify a cache directory
});

await stagehand.init();
const page = stagehand.context.pages()[0];

await page.goto("https://browserbase.github.io/stagehand-eval-sites/sites/iframe-same-proc-scroll/");

// First run: uses LLM inference and caches
// Subsequent runs: reuses cached action
await stagehand.act("scroll to the bottom of the iframe");

// Variables work with caching too
await stagehand.act("fill the username field with %username%", {
  variables: {
    username: "fakeUsername",
  },
});

​
Caching with agent()

Cache agent actions (including Computer Use Agent actions) the same way - just specify a cacheDir. The cache key is automatically generated based on the instruction, start URL, agent execution options, and agent configuration - subsequent runs with the same parameters will reuse cached actions. 
import { Stagehand } from "@browserbasehq/stagehand";

const stagehand = new Stagehand({
  env: "BROWSERBASE",
  cacheDir: "agent-cache", // Specify a cache directory
});

await stagehand.init();
const page = stagehand.context.pages()[0];

await page.goto("https://browserbase.github.io/stagehand-eval-sites/sites/drag-drop/");

const agent = stagehand.agent({
  mode: "cua",
  model: {
    modelName: "google/gemini-2.5-computer-use-preview-10-2025",
    apiKey: process.env.GOOGLE_GENERATIVE_AI_API_KEY
  },
  systemPrompt: "You are a helpful assistant that can use a web browser.",
});

await page.goto("https://play2048.co/");

// First run: uses LLM inference and caches
// Subsequent runs: reuses cached actions
const result = await agent.execute({
  instruction: "play a gane of 2048",
  maxSteps: 20,
});

console.log(JSON.stringify(result, null, 2));

​
Cache Directory Organization

You can organize your caches by using different directory names for different workflows: 
// Separate caches for different parts of your automation
const loginStagehand = new Stagehand({
  env: "BROWSERBASE",
  cacheDir: "cache/login-flow"
});

const checkoutStagehand = new Stagehand({
  env: "BROWSERBASE",
  cacheDir: "cache/checkout-flow"
});

const dataExtractionStagehand = new Stagehand({
  env: "BROWSERBASE",
  cacheDir: "cache/data-extraction"
});

​
Best Practices

Organize caches by workflow or feature for easier management:
// Good: descriptive cache names
cacheDir: "cache/login-actions"
cacheDir: "cache/search-actions"
cacheDir: "cache/form-submissions"

// Avoid: generic cache names
cacheDir: "cache"
cacheDir: "my-cache"
If the website structure changes significantly, clear your cache directory to force fresh inference:
rm -rf cache/login-actions
Or programmatically:
import { rmSync } from 'fs';

// Clear cache before running if needed
if (shouldClearCache) {
  rmSync('cache/login-actions', { recursive: true, force: true });
}

const stagehand = new Stagehand({
  env: "BROWSERBASE",
  cacheDir: "cache/login-actions"
});
Consider committing your cache directory to version control for consistent behavior across environments:
# .gitignore
# Don't ignore cache directories
!cache/
This ensures your CI/CD pipelines use the same cached actions without needing to run inference on first execution.