v3-mcp-optimization

v3 MCP Optimization Skill

Catalog of six concrete optimization patterns for EVOKORE-MCP's ProxyManager in v3. Each pattern has a measured-or-expected gain, a trigger condition, a TypeScript snippet, and an anti-pattern note. Apply these only when a profiler or benchmark identifies the corresponding bottleneck — premature application of (3) or (5) can regress cold-start.

Trigger

Use this skill when:

• Tool dispatch shows up as hot in a flame graph (indicates pattern 1)
• Child server boot dominates session startup time (indicates patterns 4 and 6)
• Repeated tool-schema lookups allocate heavily (indicates patterns 2 and 5)
• Multi-tool workflows show N× round-trip overhead (indicates pattern 3)

Pattern Catalog

1. O(1) Hash-Map Tool Lookup

Replace the linear tools.find(t => t.name === name) scan with a Map<string, Tool> built at registration time. With ~300 tools across proxied servers, this moves dispatch from O(n) to O(1).

Expected gain: 20–60× speedup on dispatch for aggregators with 100+ tools.

// Before
const tool = this.tools.find(t => t.name === name); // O(n)

// After
private toolIndex: Map<string, Tool> = new Map();
registerTool(tool: Tool) {
  this.toolIndex.set(tool.name, tool);
}
dispatch(name: string) {
  const tool = this.toolIndex.get(name); // O(1)
  if (!tool) throw new Error(`Unknown tool: ${name}`);
  return tool;
}

2. 3-Tier Cache (L1 in-memory, L2 filesystem, L3 remote registry)

Skill and tool metadata flows through a three-level cache. L1 is a Map with LRU eviction; L2 is an on-disk JSON cache under ~/.evokore/cache/; L3 is the remote registry fetch. Hits cascade downward and fill upward.

Expected gain: cold-start ~40% faster on warm L2, ~95% faster on warm L1.

async getSkill(name: string): Promise<Skill> {
  // L1
  const hit1 = this.l1.get(name);
  if (hit1 && !this.isStale(hit1)) return hit1;

  // L2
  const hit2 = await this.l2ReadJson(`skills/${name}.json`);
  if (hit2 && !this.isStale(hit2)) {
    this.l1.set(name, hit2);
    return hit2;
  }

  // L3
  const fresh = await this.registry.fetchSkill(name);
  await this.l2WriteJson(`skills/${name}.json`, fresh);
  this.l1.set(name, fresh);
  return fresh;
}