Progressive Disclosure
tools/list returns names + 60-char descriptions only. Full schemas fetched on demand and cached. 89% fewer tokens at discovery.
Standard MCP has two token bloat problems.
Tool-definition bloat. Every tool’s full JSON schema loads into context at startup — even tools the model never uses. With 10 tools you’re paying 850+ tokens before any work happens. With 50 tools across enterprise servers, thousands.
Tool-result bloat. Large outputs (file reads, search results, API responses) route through LLM context unfiltered. One 50 KB file read can destroy your context budget.
Delta-MCP fixes both.
| Standard MCP | Delta-MCP | |
|---|---|---|
| 6-tool server init | 943 tokens | 118 tokens |
| 20-tool server init | ~3 600 tokens | 378 tokens |
| Definition overhead (1/6 tools used) | 941 tokens upfront | 229 tokens on-demand |
| Tool-selection accuracy (Opus 4.5) | 79.5% | 88.1% |
| Compact-JSON wire reduction | — | −18.7% |
Progressive Disclosure
tools/list returns names + 60-char descriptions only. Full schemas fetched on demand and cached. 89% fewer tokens at discovery.
Result Handler
Large outputs truncated, paginated, or summarized automatically. The model never sees more than it needs.
Compact Wire Encoding
Negotiated at initialize. 18% smaller payloads. Auto-fallback to standard JSON for unaware clients.
OAuth 2.1
Resource-server only. Stateless JWT validation + RFC 9728 PRM discovery. One config option.