Multi-container Firefox browser for security testing — MCP + A2A server
HackBrowser MCP demonstrates moderate quality with good naming conventions and comprehensive schemas, but significant gaps in description detail, output documentation, and error handling guidance. 28 tools are explicitly defined with Zod schemas and descriptions. Tool names follow verb_noun patterns (browser_launch, container_setup, navigate, element_click, etc.) which aids LLM parsing. However, descriptions vary widely in completeness, some are specific and actionable (e.g., 'Launch Firefox with managed profile and container support'), while others are vague or lack critical context (e.g., 'Wait for an element, URL, or request to appear' lacks explanation of what 'waiting' means or what happens on timeout). Parameter descriptions are generally present but often too brief (avg ~50 chars vs. the 72-char baseline). Critically, the tools lack documented output schemas, while the code shows ToolResult structures returning JSON or text, the MCP tool definitions do not expose what fields clients should expect, forcing LLMs to infer structure. Error handling is minimal, tools log failures but provide no recovery guidance. Missing are actionable error classifications, partial-failure reporting for batch operations, and timeout handling.
Compare access levels between two containers (IDOR/privilege escalation detection).
Generate access matrix: which containers can access which endpoints.
Close the browser and auto-export HAR to disk.
Launch Firefox with managed profile and container support. Must be called before other tools.
Get current browser status: protocol, containers, tabs, captured requests.
Clear all captured requests.
Export captured HTTP requests as HAR file.
No documented output schemas in MCP tool definitions. Tools return ToolResult with content arrays, but LLMs have no visibility into what fields/structure to expect from responses (e.g., what does capture_list return, array of request objects with which fields?). This forces LLMs to infer structure and wastes context parsing ambiguous responses.
Descriptions lack actionable recovery guidance and error classifications. Tools like wait_for, test_injection, and crawl have no documented error behavior (what happens on timeout? Does inject succeed if one payload succeeds? Does crawl fail if a page fails to load?). LLMs cannot plan recovery without explicit error documentation.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 53 | 2026-07-28+ | v2 |
List all captured HTTP requests with filters.
Load captured requests from a previous HAR export.
List all containers with their roles, auth status, and tab count.
Perform login for a container. Either programmatic (fill form + submit) or manual (opens login page, waits for user).
Create and configure containers with roles and credentials. Max 4 containers. Each container has isolated cookies, storage, and sessions.
Crawl a website from a container and capture all requests.
Click an element in a container tab by CSS selector.
Read text content from an element.
Type text into an input field.
Extract API endpoints from captured requests.
Send a request to an extracted endpoint.
Find potential injection points (SQLi, XSS, SSRF, SSTI, CMDi, LFI, open redirect, mass assignment, IDOR) in captured requests.
Navigate to a URL in a specific container's tab.
Generate comprehensive security report from findings.
Close a tab.
List all tabs in a container.
Open a new tab in a container.
Test an endpoint for CSRF vulnerability.
Test an endpoint for SQL injection vulnerability.
Test an endpoint for rate limiting protection.
Wait for an element, URL, or request to appear.
Many parameter descriptions are too brief (<30 chars) and lack format/constraint guidance. E.g., wait_for's 'value' param says 'Selector, URL pattern, or request pattern' but does not explain what a valid 'request pattern' looks like. crawl's 'scope' enum values ('same-domain' vs 'same-origin') lack explanation of the difference. LLMs often misuse ambiguous parameters.
No pagination support or result limits documented. Tools like capture_list accept optional 'limit' param (default 100) but do not explain what happens if there are 10,000 requests, does it return all or cap silently? No cursor/offset return fields for continuation. Analysis tools (access_matrix, injection_find) risk returning massive result sets that exhaust context windows.
No dry-run or confirmation support for destructive/high-risk operations. container_setup with credentials, browser_close with HAR export, capture_clear, test_ratelimit (generates high request volume), and injection/CSRF tests modify state or trigger security scans with no rollback option. Agents cannot preview or confirm before executing.
Security testing tools (test_injection, test_csrf, test_ratelimit, crawl) lack rate limiting, timeout enforcement, and resource bounds documentation. No guidance on max payload size, max test duration, or request throttling. Agents could inadvertently launch denial-of-service attacks against target applications or themselves.
No tool annotations (readOnlyHint, destructiveHint, idempotentHint) present in schema. Tests, writes, and destructive operations should be explicitly marked so agents know which tools are safe to retry or parallelize. E.g., browser_launch should be marked idempotent-to-first-success, while capture_clear should be marked destructive.
Parameter names could be more specific and self-documenting. Many use generic terms: 'value' (in wait_for), 'type' (in container_setup credential type), 'fields' (in multiple tools). Shadowing across tools creates ambiguity. E.g., endpoint_request's 'headers' object lacks field-level description of whether standard HTTP headers (Content-Type, Authorization) are auto-managed or manual.
No clear dependency chain documented between tools. LLMs must infer that browser_launch must be called first, that container_setup should precede navigation, that endpoint_extract requires prior crawl/capture. These dependencies should be documented in descriptions with explicit ordering hints.
Some tool names could be more specific. 'crawl' is ambiguous, does it mean web crawler, code crawler, or configuration crawler? 'endpoint_request' vs 'send_request' or 'http_request' would be clearer. 'test_*' tools could use 'check_*' or 'scan_*' to distinguish them from assertion/validation tests.