Skip to main content

​
Dynamic Tool Discovery

When an MCP server exposes many tools, the standard tools/list call returns the full JSON schema for every single tool upfront. For servers with 50 or more tools, this alone can cost 10,000–25,000 tokens before the LLM has done any real work. Dynamic Tool Discovery solves this by replacing that upfront dump with an on-demand discovery pattern, cutting token usage by up to 95%.

​
The Problem

Each tool schema averages 200–500 tokens. A server with 100 tools therefore consumes up to 50,000 tokens just to initialize — a significant portion of most models’ context windows. This overhead:
  • Increases inference cost on every conversation turn
  • Leaves less room for actual task context and conversation history
  • Slows down responses due to larger prompt sizes

​
How It Works

When Dynamic Tool Discovery is enabled, HasMCP wraps the full toolset behind three standardized discovery tools that the LLM interacts with instead:
ToolPurpose
searchToolsSearch for relevant tools by keyword or regex pattern
getToolDefinitionRetrieve the full schema for a specific tool, on demand
useToolExecute any tool by name and arguments
The LLM only ever loads the schemas it actually needs, keeping the context lean throughout the session.

​
Hybrid Search Engine

searchTools is backed by a hybrid search engine combining two algorithms:
  • BM25 ranking — probabilistic relevance matching with smart tokenization that treats getUser, get_user, and get-user as equivalent. Handles camelCase, snake_case, and kebab-case conventions automatically.
  • Regex matching — enables precise pattern searches like ^stripe.*(charge|refund)$ with case-insensitive enforcement.
BM25 matches are ranked first; regex matches follow. This gives the LLM both fuzzy relevance search and exact pattern targeting in a single call.

​
Key Benefits

  • Up to 95% token reduction for large toolsets
  • Tools can be added or removed with zero client-side reloads
  • Full schema detail is available on demand — nothing is lost, just deferred
  • Enables “Mega-Servers” with hundreds of tools that would otherwise be impractical

​
Comparison with Similar Approaches

Dynamic Tool Discovery operates at the MCP protocol level, within a single server. This is different from:
  • Claude’s tool-search-tool — client-specific, not portable across MCP clients
  • Docker’s dynamic-mcp — server discovery (finding servers), not tool discovery within a server
HasMCP’s approach is client-agnostic and works with any MCP-compatible AI tool.