Four tools with clear action verbs and detailed descriptions (100-200 chars each). All have input schemas with types and descriptions. agent_browser_act is particularly well-designed with enum constraints (action, robust) and parameter relationships documented. However, output schemas are not explicitly documented in the source code, only inferred from descriptions. No error handling guidance or recovery patterns visible. Missing tool annotations (readOnlyHint/destructiveHint). Risk classifications are present but not formally integrated into tool metadata.

• Document output schemas explicitly. For agent_browser_visit and agent_browser_act, define the accessibility-tree structure: { nodes: [{id, role, name, value, children}], ...}. Include field types and whether results are paginated.
• Add error recovery hints to descriptions. E.g., agent_browser_ensure_running: 'If launch fails, verify host-helper is running (npm run host-helper:status) and Chrome is not already open.'
• Expose tool annotations via MCP. Register agent_browser_status and agent_browser_visit with readOnlyHint=true; agent_browser_ensure_running and agent_browser_act with destructiveHint=true (or idempotentHint for ensure_running).
• Clarify actionClass validation. Document which domains are 'trusted' (e.g., internal company domains) and what happens when an untrusted domain attempts a sensitive action (e.g., credential_entry). Provide an error message template.
• Add result-size guidance. Specify max accessibility-tree depth, node count, or response size. If truncation occurs, document how the LLM can request a subtree or navigate incrementally.