Comet MCP has six tools with complete JSON Schema definitions and descriptions present in src/index.ts. However, the server exhibits significant quality gaps across the definition rubric. Tool names lack clear action verbs (e.g., 'comet_ask' is vague, 'send_prompt' or 'query_comet' would be clearer). Parameter descriptions are sparse or missing entirely: comet_ask has good descriptions for its three parameters, but comet_connect, comet_poll, comet_stop, and comet_screenshot have empty input schemas with no properties to describe. Output schemas are entirely undocumented, the LLM cannot infer what these tools return, forcing it to guess at downstream usage. Error handling is not visible in the provided source, and no recovery guidance is documented. The tool definitions are statically declared and directly visible, so they are not inferred, avoiding the 50-cap penalty. However, sparse parameter coverage and missing output documentation significantly depress the score.

• Document the complete output schema for every tool. Example for comet_screenshot: 'Returns {"type": "text", "content": {"imageBase64": "...", "width": 1920, "height": 1080, "format": "png"}}'. This enables the LLM to understand what data is available and how to use it downstream.
• Rename tools with clearer action verbs: comet_ask → send_comet_prompt (or query_comet); comet_poll → get_agent_status; comet_mode → set_search_mode; comet_connect → start_comet_session. Per pattern:tool, verb_noun naming lets LLMs parse intent from the name alone.
• Add bounds to numeric parameters. Example: timeout parameter: 'Max wait time in ms (default: 15000 = 15s, valid range: 1000 - 300000)'. This prevents absurd parameter values.
• Document error scenarios and recovery paths. Example for comet_ask: 'If the request times out, call comet_poll() to check if the agent is still running. If comet_poll() indicates idle, try again with a longer timeout or simpler prompt.' This follows pattern:recovery-guide.
• Clarify parameter relationships and dependencies. Example: explain in comet_mode description that changing the mode affects the next comet_ask call, and does NOT retroactively change the mode of previous responses. Document whether mode persists across comet_connect calls.
• For parameters like newChat in comet_ask, add an explicit description of the pre-condition and side effect: 'If true, closes all existing tabs and starts a fresh conversation at perplexity.ai. Use when the agent got stuck or needs a clean slate. Default: false (continues existing conversation).' This reduces ambiguity.