Inject MCP tools into fighter conversations with lazy loading

[humancto]

Problem

Even once MCP servers are spawned and tools discovered, fighters will not see them unless tools are injected into the LLM tool list.

Token Bloat Problem

This is the #1 complaint about MCP in 2026. 10 servers x 5 tools = 50 tool definitions consuming 30K+ tokens before the user says anything. OpenClaw has no solution. Perplexity CTO publicly abandoned MCP over this.

Proposed Solution

Lazy tool loading (like Claude Code Tool Search):

• On first message, inject a discover_tools meta-tool that lists available MCP servers + tool summaries
• Fighter calls discover_tools("gmail") to load Gmail tools into context on-demand
• Only actively-used tools consume context window tokens
• Estimated 40-50% token reduction vs naive loading

Implementation

1. Convert MCP tool schemas to OpenAI-compatible function definitions
2. When LLM calls an MCP tool, route to the correct server via tools/call
3. Return result back to the LLM conversation
4. Cache tool schemas to avoid repeated tools/list calls