browser-use 0.6.0 → 0.6.1

package/README.md

@@ -23,7 +23,7 @@
 ## ✨ Features
 
 - 🤖 **Autonomous Browser Control** — AI-driven navigation, clicking, typing, form filling, scrolling, and tab management
-- 🧠 **10+ LLM Providers** — OpenAI, Anthropic, Google Gemini, Azure, AWS Bedrock, Groq, Ollama, DeepSeek, OpenRouter, Mistral, Cerebras, and custom providers
+- 🧠 **15+ LLM Providers & Adapters** — OpenAI, Anthropic, Google Gemini, Azure, AWS Bedrock, Groq, Ollama, DeepSeek, OpenRouter, Mistral, Cerebras, Browser Use, LiteLLM, OCI Raw, Vercel, and custom providers
 - 👁️ **Vision Support** — Screenshot-based understanding for visual web interactions
 - 🔧 **45+ Built-in Actions** — Navigation, element interaction, scrolling, forms, tabs, content extraction, file I/O, and more
 - 🧩 **Custom Actions** — Extensible registry with Zod schema validation, domain restrictions, and page filters
@@ -112,7 +112,7 @@ npx browser-use --mcp
 | **BrowserSession** | Playwright wrapper — browser lifecycle, tab management, screenshots    |
 | **DomService**     | Extracts interactive elements with indexed mapping for LLM consumption |
 | **MessageManager** | Manages LLM conversation history with token optimization               |
-| **LLM Providers**  | Unified `BaseChatModel` interface across 10+ providers                 |
+| **LLM Providers**  | Unified `BaseChatModel` interface across 15+ providers and adapters    |
 
 ### How It Works
 
@@ -125,19 +125,23 @@ npx browser-use --mcp
 
 ## 🔌 LLM Providers
 
-| Provider          | Import                       | Vision | Notes                                         |
-| ----------------- | ---------------------------- | ------ | --------------------------------------------- |
-| **OpenAI**        | `browser-use/llm/openai`     | ✅     | Default provider, reasoning models (o1/o3/o4) |
-| **Anthropic**     | `browser-use/llm/anthropic`  | ✅     | Prompt caching support                        |
-| **Google Gemini** | `browser-use/llm/google`     | ✅     | Extended thinking support                     |
-| **Azure OpenAI**  | `browser-use/llm/azure`      | ✅     | Enterprise deployment                         |
-| **AWS Bedrock**   | `browser-use/llm/aws`        | ✅     | Claude via AWS                                |
-| **Groq**          | `browser-use/llm/groq`       | ❌     | Fastest inference                             |
-| **Ollama**        | `browser-use/llm/ollama`     | ❌     | Local/self-hosted models                      |
-| **DeepSeek**      | `browser-use/llm/deepseek`   | ❌     | Cost-effective                                |
-| **OpenRouter**    | `browser-use/llm/openrouter` | Varies | Multi-model routing                           |
-| **Mistral**       | `browser-use/llm/mistral`    | Varies | Mistral models                                |
-| **Cerebras**      | `browser-use/llm/cerebras`   | ❌     | Fast inference                                |
+| Provider          | Import                        | Vision | Notes                                         |
+| ----------------- | ----------------------------- | ------ | --------------------------------------------- |
+| **OpenAI**        | `browser-use/llm/openai`      | ✅     | Default provider, reasoning models (o1/o3/o4) |
+| **Anthropic**     | `browser-use/llm/anthropic`   | ✅     | Prompt caching support                        |
+| **Google Gemini** | `browser-use/llm/google`      | ✅     | Extended thinking support                     |
+| **Azure OpenAI**  | `browser-use/llm/azure`       | ✅     | Enterprise deployment                         |
+| **AWS Bedrock**   | `browser-use/llm/aws`         | ✅     | Claude via AWS                                |
+| **Groq**          | `browser-use/llm/groq`        | ❌     | Fastest inference                             |
+| **Ollama**        | `browser-use/llm/ollama`      | ❌     | Local/self-hosted models                      |
+| **DeepSeek**      | `browser-use/llm/deepseek`    | ❌     | Cost-effective                                |
+| **OpenRouter**    | `browser-use/llm/openrouter`  | Varies | Multi-model routing                           |
+| **Mistral**       | `browser-use/llm/mistral`     | Varies | Mistral models                                |
+| **Cerebras**      | `browser-use/llm/cerebras`    | ❌     | Fast inference                                |
+| **Browser Use**   | `browser-use/llm/browser-use` | Varies | Hosted Browser Use LLM                        |
+| **LiteLLM**       | `browser-use/llm/litellm`     | Varies | OpenAI-compatible LiteLLM gateway             |
+| **OCI Raw**       | `browser-use/llm/oci-raw`     | Varies | Oracle Cloud Generative AI                    |
+| **Vercel**        | `browser-use/llm/vercel`      | Varies | Vercel AI Gateway / routed models             |
 
 <details>
 <summary>Provider examples</summary>
@@ -210,6 +214,7 @@ const agent = new Agent({
 ### Custom Actions
 
 ```typescript
+import fs from 'node:fs';
 import { Controller, ActionResult } from 'browser-use';
 import { z } from 'zod';
 
@@ -345,7 +350,7 @@ Add to your Claude Desktop config (`~/Library/Application Support/Claude/claude_
 }
 ```
 
-Available MCP tools: `browser_run_task`, `browser_navigate`, `browser_click`, `browser_type`, `browser_scroll`, `browser_get_state`, `browser_extract`, `browser_screenshot`, `browser_close`.
+Core MCP tools include `retry_with_browser_use_agent`, `browser_navigate`, `browser_click`, `browser_type`, `browser_get_state`, `browser_extract_content`, `browser_scroll`, `browser_go_back`, `browser_list_tabs`, `browser_switch_tab`, `browser_close_tab`, `browser_list_sessions`, `browser_close_session`, and `browser_close_all`. The server also exposes registered controller actions as additional MCP tools.
 
 > See [MCP Server Guide](./docs/MCP_SERVER.md) for more details.
 
@@ -423,4 +428,4 @@ pnpm exec tsx examples/simple-search.ts
 
 ## 📄 License
 
-[MIT](./LICENSE) © Web LLM
+[MIT](./LICENSE)

package/dist/agent/service.js

@@ -34,7 +34,7 @@ import { AgentTelemetryEvent } from '../telemetry/views.js';
 import { TokenCost } from '../tokens/service.js';
 import { construct_judge_messages, construct_simple_judge_messages, } from './judge.js';
 import { CloudSkillService, MissingCookieException, build_skill_parameters_schema, get_skill_slug, } from '../skills/index.js';
-loadEnv();
+loadEnv({ quiet: true });
 const logger = createLogger('browser_use.agent');
 const URL_PATTERN = /https?:\/\/[^\s<>"']+|www\.[^\s<>"']+|[^\s<>"']+\.[a-z]{2,}(?:\/[^\s<>"']*)?/gi;
 export const log_response = (response, registry, logInstance = logger) => {
@@ -3886,7 +3886,16 @@ export class Agent {
                 {});
             const paramsResult = actionInfo.paramSchema.safeParse(rawParams);
             if (!paramsResult.success) {
-                throw new Error(`Invalid parameters for action '${requestedActionName}': ${paramsResult.error.message}`);
+                // Surface a human-readable issue list (zod v4 `prettifyError`) plus
+                // a corrective hint, rather than the default JSON dump of `.issues`.
+                // This Error propagates → `_handle_step_error` writes it into
+                // `state.last_result` → `create_state_messages` injects it into the
+                // next LLM turn, so the model knows exactly what shape it got wrong.
+                const pretty = z.prettifyError(paramsResult.error);
+                const sentParams = JSON.stringify(rawParams);
+                throw new Error(`Schema validation failed for action '${requestedActionName}'. ` +
+                    `You sent: ${sentParams}. Issues:\n${pretty}\n` +
+                    `Please retry with parameters matching the action's schema exactly.`);
             }
             normalizedActions.push(new modelForStep({
                 [actionName]: paramsResult.data,

package/dist/cli.js

@@ -33,7 +33,7 @@ import { setupLogging } from './logging-config.js';
 import { get_tunnel_manager } from './skill-cli/tunnel.js';
 import { DeviceAuthClient, save_cloud_api_token } from './sync/auth.js';
 import dotenv from 'dotenv';
-dotenv.config();
+dotenv.config({ quiet: true });
 const require = createRequire(import.meta.url);
 const CLI_PROVIDER_ALIASES = {
     openai: 'openai',

package/dist/config.js

@@ -4,7 +4,7 @@ import path from 'node:path';
 import { randomUUID } from 'node:crypto';
 import { config as loadEnv } from 'dotenv';
 import { createLogger } from './logging-config.js';
-loadEnv();
+loadEnv({ quiet: true });
 const logger = createLogger('browser_use.config');
 const expand_user = (value) => value.replace(/^~(?=$|\/|\\)/, os.homedir());
 const resolve_path = (value) => path.resolve(expand_user(value));

package/dist/controller/registry/views.d.ts

@@ -4,6 +4,7 @@ export type ActionHandler = (...args: any[]) => Promise<unknown> | unknown;
 type BrowserSession = unknown;
 type BaseChatModel = unknown;
 type FileSystem = unknown;
+export declare function renderParamsJsonSchema(schema: ZodTypeAny, skipKeys: Set<string>): Record<string, unknown>;
 export declare class RegisteredAction {
     readonly name: string;
     readonly description: string;
@@ -13,6 +14,7 @@ export declare class RegisteredAction {
     readonly pageFilter: ((page: Page) => boolean) | null;
     readonly terminates_sequence: boolean;
     constructor(name: string, description: string, handler: ActionHandler, paramSchema: ZodTypeAny, domains?: string[] | null, pageFilter?: ((page: Page) => boolean) | null, terminates_sequence?: boolean);
+    getPromptJsonSchema(): Record<string, unknown>;
     promptDescription(): string;
 }
 export declare class ActionModel {

package/dist/controller/registry/views.js

@@ -15,6 +15,38 @@ const getPageUrl = (page) => {
     }
     return candidate ?? '';
 };
+// Render an action's param schema as compact JSON Schema for the LLM prompt.
+// Replaces a prior raw dump of zod's private `_def` AST, which leaked
+// internal keys like `innerType`/`defaultValue` and confused the LLM into
+// copying default booleans into numeric fields (see scroll.num_pages bug).
+export function renderParamsJsonSchema(schema, skipKeys) {
+    // `io: 'input'` makes zod render the *input* shape (what the LLM is
+    // expected to provide). Without it, fields with `.default(...)` get marked
+    // as required in the JSON Schema (because the parsed *output* always has
+    // them), which misleads the model — e.g. scroll.num_pages, done.success.
+    const raw = z.toJSONSchema(schema, {
+        io: 'input',
+        unrepresentable: 'any',
+    });
+    // Strip dialect noise the LLM doesn't need.
+    delete raw.$schema;
+    const properties = raw.properties ?? {};
+    const filteredProps = {};
+    for (const [key, value] of Object.entries(properties)) {
+        if (skipKeys.has(key)) {
+            continue;
+        }
+        filteredProps[key] = value;
+    }
+    raw.properties = filteredProps;
+    if (Array.isArray(raw.required)) {
+        raw.required = raw.required.filter((key) => typeof key === 'string' && !skipKeys.has(key));
+        if (raw.required.length === 0) {
+            delete raw.required;
+        }
+    }
+    return raw;
+}
 export class RegisteredAction {
     name;
     description;
@@ -32,10 +64,12 @@ export class RegisteredAction {
         this.pageFilter = pageFilter;
         this.terminates_sequence = terminates_sequence;
     }
-    promptDescription() {
+    // Returns the JSON Schema rendered for the LLM prompt, with the same
+    // skipKeys logic applied as in `promptDescription`. Exposed so tooling
+    // (e.g. scripts/dump-schema.ts) can exercise the exact code path the
+    // model sees.
+    getPromptJsonSchema() {
         const skipKeys = new Set(['title']);
-        let description = `${this.description}: \n`;
-        description += `{${this.name}: `;
         const schemaShape = (this.paramSchema instanceof z.ZodObject && this.paramSchema.shape) ||
             ('shape' in this.paramSchema ? this.paramSchema.shape : null);
         const hideStructuredDoneSuccess = Boolean(this.name === 'done' &&
@@ -46,26 +80,19 @@ export class RegisteredAction {
         if (hideStructuredDoneSuccess) {
             skipKeys.add('success');
         }
-        const hideExtractOutputSchema = Boolean(this.name === 'extract_structured_data' &&
+        const hideExtractOutputSchema = Boolean((this.name === 'extract_structured_data' || this.name === 'extract') &&
             schemaShape &&
             typeof schemaShape === 'object' &&
             Object.prototype.hasOwnProperty.call(schemaShape, 'output_schema'));
         if (hideExtractOutputSchema) {
             skipKeys.add('output_schema');
         }
-        if (schemaShape) {
-            const props = Object.fromEntries(Object.entries(schemaShape)
-                .filter(([key]) => !skipKeys.has(key))
-                .map(([key, value]) => {
-                const entries = value instanceof z.ZodType ? value._def : value;
-                const cleanEntries = Object.fromEntries(Object.entries(entries).filter(([propKey]) => !skipKeys.has(propKey)));
-                return [key, cleanEntries];
-            }));
-            description += JSON.stringify(props);
-        }
-        else {
-            description += '{}';
-        }
+        return renderParamsJsonSchema(this.paramSchema, skipKeys);
+    }
+    promptDescription() {
+        let description = `${this.description}: \n`;
+        description += `{${this.name}: `;
+        description += JSON.stringify(this.getPromptJsonSchema());
         description += '}';
         return description;
     }

package/dist/observability.js

@@ -1,7 +1,7 @@
 import { createRequire } from 'node:module';
 import { config as loadEnv } from 'dotenv';
 import { createLogger } from './logging-config.js';
-loadEnv();
+loadEnv({ quiet: true });
 const require = createRequire(import.meta.url);
 const logger = createLogger('browser_use.observability');
 let lmnrObserve = null;

package/dist/utils.js

@@ -10,7 +10,7 @@ import { fileURLToPath } from 'node:url';
 import { config as loadEnv } from 'dotenv';
 import * as minimatchModule from 'minimatch';
 import { createLogger } from './logging-config.js';
-loadEnv();
+loadEnv({ quiet: true });
 const logger = createLogger('browser_use.utils');
 let _exiting = false;
 const minimatch = (minimatchModule.minimatch ??

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "browser-use",
-  "version": "0.6.0",
+  "version": "0.6.1",
   "description": "A TypeScript-first library for programmatic browser control, designed for building AI-powered web agents.",
   "main": "./dist/index.js",
   "module": "./dist/index.js",
@@ -264,6 +264,11 @@
   "engines": {
     "node": ">=18.0.0"
   },
+  "config": {
+    "commitizen": {
+      "path": "cz-conventional-changelog"
+    }
+  },
   "keywords": [
     "browser",
     "use",
@@ -285,7 +290,7 @@
     "@google/genai": "^1.40.0",
     "@modelcontextprotocol/sdk": "^1.27.1",
     "adm-zip": "^0.5.16",
-    "axios": "^1.13.4",
+    "axios": "^1.16.0",
     "canvas": "^3.2.1",
     "dotenv": "^17.2.4",
     "eventemitter3": "^5.0.4",
@@ -318,6 +323,7 @@
     "@typescript-eslint/parser": "^8.54.0",
     "@vitest/coverage-v8": "^4.0.18",
     "commitizen": "^4.3.1",
+    "cz-conventional-changelog": "^3.3.0",
     "eslint": "^9.39.2",
     "eslint-config-prettier": "^10.1.8",
     "eslint-plugin-import": "^2.32.0",
@@ -326,14 +332,9 @@
     "prettier": "^3.8.1",
     "tsx": "^4.21.0",
     "typescript": "^5.9.3",
-    "vite": "^7.3.1",
+    "vite": "^7.3.2",
     "vitest": "^4.0.18"
   },
-  "config": {
-    "commitizen": {
-      "path": "cz-conventional-changelog"
-    }
-  },
   "scripts": {
     "build": "node scripts/clean-dist.mjs && tsc && node scripts/copy-dom-tree.mjs",
     "build:watch": "tsc --watch --preserveWatchOutput",
@@ -349,12 +350,12 @@
     "test:watch": "vitest --watch",
     "test:pack": "node scripts/smoke-pack.mjs",
     "check": "pnpm lint && pnpm typecheck && pnpm typecheck:test && pnpm test:unit && pnpm test:integration && pnpm test:e2e && pnpm test:pack",
+    "commit": "cz",
     "typecheck": "tsc --noEmit",
     "typecheck:test": "tsc -p tsconfig.test.json --noEmit",
     "format": "prettier --write \"src/**/*.ts\" \"test/**/*.ts\"",
     "format:check": "prettier --check \"src/**/*.ts\" \"test/**/*.ts\"",
     "prettier": "prettier --write \"src/**/*.ts\" \"test/**/*.ts\"",
-    "commit": "pnpm exec git-cz",
     "postinstall": "playwright install chromium",
     "postinstall:ci": "playwright install --with-deps chromium"
   }