AI debugger, inspector, and verification toolkit for local-first browser development workflows. Streams console logs, network errors, and exceptions to Claude Code, Copilot, Cursor, or any MCP-compatible assistant.
Critical failures across all dimensions. Tools are defined only in test files with minimal detail. The three tools (observe, interact, configure) have extremely brief descriptions (10-19 chars), no visible input schemas despite being listed with partial parameter info, no output schemas documented, and no evidence of error handling guidance. The server is a custom STDIO-based browser control tool, but tool definitions lack the rigor required for production LLM agents. Schema information appears only in a test file (handler_tools_call_usage_test.go), not in the canonical tool registration code, making it inferred rather than directly defined.
Configure browser and server settings
Interact with the browser: navigate, click, input, scroll, and execute commands
Observe browser state, console logs, errors, and network activity
Tool definitions inferred from test file only, no canonical registration visible. Schema and description sourced from handler_tools_call_usage_test.go rather than main tool registration code.
All three tools have descriptions under 20 characters. 'Observe browser state...' (19 chars for observe), 'Interact with the browser...' (13 chars for interact), 'Configure browser...' (10 chars for configure). Descriptions must be 10 - 1024 chars and answer WHAT, WHEN, and prerequisites.
No input schemas visible in source for any tool. The handler_tools_call_usage_test.go shows parameter info for 'observe' (what, correlation_id) but no JSON Schema structure, types, or validation rules are present in canonical code.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-22 | F | 26 | 2026-07-28+ | v2 |
| 2026-03-09 | F | 0 | - | v1 |
No output schemas documented. Tools return unspecified results. LLMs cannot plan downstream tool calls or extract required fields (e.g., command_result field for interact's return value is undefined). Baseline: 100% of A+ tools have documented return types.
Missing error handling and recovery guidance. No evidence of pattern:recovery-guide or pattern:error-classification. Tools do not indicate what the LLM should do on failure (retry? ask user? abort?). interact is destructive (WRITE risk), no confirmation or dry-run pattern visible.
No parameter descriptions visible. The 'observe' tool lists 'what' and 'correlation_id' parameters but neither has a description explaining expected values, constraints, or format. BASELINE: 100% of A+ tools have descriptions for all parameters.
Parameter 'what' in observe lacks enum constraint. Accepts free-form strings ('errors', 'logs', 'network', 'page', 'command_result', or other), invites hallucinated values. Should declare as enum: ["errors", "logs", "network", "page", "command_result"].
No composition guidance. Tool naming (interact, configure) does not indicate what specific actions are available (navigate? click? scroll? execute script?). Description is too brief to guide when to use interact vs observe. Agents cannot compose multi-step browser workflows confidently.
interact tool is marked WRITE risk but has no confirmation pattern. Destructive operations (navigate, interact with DOM) should support dry-run or explicit confirmation. No pattern:confirmation-request visible.
Naming ambiguity: 'configure' suggests system configuration but description says 'Configure browser and server settings', unclear whether this controls headless mode, viewport, user agent, or server behavior. Names must disambiguate intent.