A lazy-loading MCP proxy that cuts your agent's startup context from tens of thousands of tokens to a few hundred — it exposes upstream tool schemas only when they're actually needed.
mcp-lean is a proxy/discovery tool with 5 well-named, verb-first tools (search_tools, list_tools, load_tools, unload_tools, describe_tool). All tools have clear descriptions (60 - 180 chars) and proper input schemas with typed parameters. However, output schemas are not documented, LLMs cannot predict what search_tools or describe_tool return. Parameter descriptions are present but minimal (e.g., 'Tools to activate' lacks format guidance). No error handling guidance (what if a tool doesn't exist? what if load fails?). The tool set is cohesive and solves a real problem (context window optimization), but lacks production-grade robustness in error messaging and output documentation.
Show the full JSON schema of a single upstream tool WITHOUT activating it. Handy to inspect arguments before deciding to load.
List the names + one-line descriptions of all available upstream tools (optionally filtered to one server). Use search_tools to find by capability.
Activate one or more upstream tools by name. After loading, each appears as a first-class tool (named "server__tool") that you can call directly with its real arguments. Load only what you need.
Search across all tools from connected MCP servers by keyword. Returns one-line matches; full schemas are not loaded until you call load_tools.
Deactivate previously loaded tools to reclaim context window. Omit the argument to unload all of them.
Output schemas not documented. LLMs cannot predict what search_tools, list_tools, or describe_tool return. This forces agents to guess at response structure and risks parsing failures.
No error handling guidance. If load_tools fails (tool not found, permission denied, upstream server down), the tool provides no recovery hint. LLMs cannot self-correct.
Parameter descriptions lack format/constraint details. 'Tools to activate' does not explain the 'server/tool' or 'server/*' syntax. LLMs must infer from examples in the description text.
No pagination or result limits documented for list_tools and search_tools. If an upstream server exposes 100+ tools, these tools could return unbounded lists, exhausting context.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 69 | 2025-06-18+ | v2 |
load_tools and unload_tools lack idempotency guarantees. Calling load_tools twice with the same tool could fail on the second call if already loaded. No mention of idempotent behavior.