Loading ToolBench
AINative Browser Agent MCP Server — 6 AI-powered browser automation tools for web interaction, structured data extraction, page validation, multi-step tasks, and memory enrichment. Auto-detects ZeroLocal vs Cloud.
8 tools with complete input schemas and descriptions. Naming follows verb_noun pattern (browser_act, browser_extract, etc.). Descriptions are adequate (50-200 chars) and include credit costs. Tool annotations present (readOnlyHint, destructiveHint, idempotentHint, openWorldHint). However, output schemas are NOT documented, LLMs cannot infer what fields to expect from responses. Parameter descriptions are minimal (e.g., 'The URL of the web page to act on' lacks format/constraint details). No error handling guidance visible. No pagination support documented for batch operations. Missing recovery hints for common failure modes.
Perform an action on a web page using AI-powered browser automation. Use when you need to click, type, navigate, or otherwise interact with a page. (50 credits)
Extract structured data from up to 10 URLs in a single call and store rows in a ZeroDB table. Use when you need to scrape multiple pages at once. (75 credits per URL)
Extract content from a web page and store it in agent memory (ZeroDB). Use when you want an agent to "remember" information from a URL for future recall. (100 credits)
Queue an async browser memory enrichment task. Returns a task_id immediately — the browser visits the URL and stores content in ZeroMemory in the background. (100 credits)
Extract structured data from a web page. Use when you need to scrape, parse, or retrieve specific information from a URL. Returns data in structured JSON format. (75 credits)
Output schemas not documented. LLMs cannot infer what fields responses contain (e.g., does browser_extract return {data, metadata, error}? Is it an array or object?). This forces agents to guess and wastes tokens on trial-and-error parsing.
Parameter descriptions lack format/constraint details. 'The URL of the web page to act on' does not specify: must it be http/https? Can it be relative? What happens if the URL is invalid? LLMs need explicit constraints to avoid malformed inputs.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 64 | 2026-07-28+ | v2 |
Extract structured data from a web page and store it directly in a ZeroDB NoSQL table. Use when you need to scrape data and persist it for later querying or analysis. (100 credits)
Run a multi-step browser automation task. Use when you need to perform a complex, multi-step workflow on a web page (e.g. login and fill out a form). More capable than browser_act for complex flows. (200 credits)
Validate content or state on a web page. Use when you need to assert that specific content exists, a form was submitted, or a page reached an expected state. (25 credits)
No error handling guidance. If browser_act fails (page timeout, element not found, JavaScript error), what should the LLM do? Retry? Call a different tool? Ask the user? Silent failures block agent progress.
browser_batch_extract accepts up to 10 URLs but no pagination or result-limiting strategy documented. If extraction returns 1000+ rows per URL, response could exceed context window. No guidance on handling partial failures (e.g., 7 of 10 URLs succeed).
ZeroDB integration (browser_extract_to_table, browser_enrich_memory, browser_batch_extract) requires project_id parameter but no guidance on how to obtain it or what happens if it's invalid. Dependency on external service not documented.