The server has 4 tools with explicit Zod schema definitions visible in src/index.ts. Three tools (screenshot, pdf, metadata) are well-defined with detailed parameter schemas and descriptions. The fourth tool (run_tool) is defined only in the tool handler but lacks a visible input schema in the code, it references external Junipr API slugs without exposing them as enumerated options or validated inputs. Parameter descriptions are generally solid (72-250 chars, meeting baseline), but the pdf tool has a design issue: it requires either 'url' or 'html' but Zod's optional() allows both to be omitted, creating runtime validation overhead. Error handling is present (try/catch, API error parsing) but messages are generic ('Screenshot failed: {error}') rather than actionable recovery guidance. No tool uses readOnlyHint or idempotentHint annotations. The run_tool parameters are not visible, preventing full schema assessment.

• For run_tool: add an input schema with a 'slug' parameter that is an enum or constrained string validated against the list of 75+ available Junipr slugs. Include descriptions of common slugs (address-validator, brave-search, etc.) to aid LLM discovery.
• For pdf tool: restructure Zod schema to enforce exactly-one-of url/html using Zod.discriminatedUnion() or a custom refinement. Error message should be: 'Must provide either url or html (not both). url: valid HTTP URL; html: valid HTML string.'
• Add readOnlyHint: true to all tool definitions (screenshot, pdf, metadata, run_tool) since they perform no state mutations. This signals to LLMs that repeated calls are safe.
• Document output schemas inline in tool descriptions. Example for screenshot: 'Returns {content: [{type: "image", data: base64, mimeType: "image/png|jpeg|webp"}]}'.
• Enrich error messages with actionable recovery guidance. Example: 'Screenshot failed: timeout. Common causes: (1) URL unreachable; (2) rate limit (check JUNIPR_API_KEY quota); (3) page too slow. Retry in 5s or try a different URL.'
• For metadata tool: add optional parameters to filter extraction scope (e.g. include_og_tags, include_json_ld, include_twitter_cards) to reduce token overhead and let LLMs request only needed metadata fields.
• Add a 'dry_run' or validation parameter to pdf and screenshot tools to allow LLMs to test parameters without consuming quota. This reduces accidental errors and API cost.