All 4 tools have basic descriptions (10-50 chars) and empty input schemas. Three tools (get_stack, get_projects, get_full_context) accept no parameters and lack any schema definition. get_context_for has a single 'task' parameter with type and description, but the tool descriptions are too brief to guide LLM selection. No output schemas documented. No error handling guidance beyond authentication checks. Tools return unstructured strings rather than structured objects. No pagination, no field documentation, no recovery hints.

• Add input schemas to all tools. Even no-parameter tools should declare 'input': {'type': 'object', 'properties': {}, 'required': []}. This signals to LLMs that the tool accepts no arguments.
• Expand tool descriptions to 100-200 characters. Example: 'get_stack: Returns the developer's tech stack (languages, frameworks, tools). Call this first to understand what technologies the developer knows before suggesting tools or frameworks.'
• Document output schemas. Example for get_stack: 'Returns a markdown string with sections for Languages, Frameworks, and Tools. Each section lists items with brief descriptions.'
• Replace string error responses with structured error objects: {'error': 'not_authenticated', 'message': 'Include your MCP token in the request.', 'recovery': 'Authenticate via /login first.'}
• Add timeout (5-10s) and retry logic to get_context_for's Gemini call. Return partial results (unfiltered context) if LLM call fails, rather than a complete error.