warmplane

Next-Level Tool Calling

Most teams think tool calling costs come from execution. In practice, a huge share comes from description overhead: sending large tool schemas and metadata before useful work starts.

That is where Warmplane changes the game.

The Core Idea

Keep MCP richness in the backend, but present agents with a compact, lazy, stable facade:

• GET /v1/capabilities
• GET /v1/capabilities/:id
• POST /v1/tools/call
• plus compact resources/prompts surfaces

This moves agent interaction from “dump everything up front” to “index first, detail on demand.”

What We Measured (Not Hypothetical)

From eval/token-efficiency/output/report.md:

GitHub Copilot MCP (authenticated)

• Discovery scenario:
  • Raw: 54,715 tokens
  • Facade: 2,386 tokens
  • Savings: 52,329 tokens (95.6%)
• 5-turn tool loop:
  • Raw: 260,005
  • Facade: 11,173
  • Savings: 248,832 (95.7%)
• 10-turn mixed loop:
  • Raw: 547,150
  • Facade: 22,895
  • Savings: 524,255 (95.8%)

Public filesystem MCP (no-auth control)

• Discovery:
  • Raw: 2,552
  • Facade: 1,066
  • Savings: 1,486 (58.2%)
• 5-turn tool loop:
  • Raw: 12,760
  • Facade: 5,349
  • Savings: 7,411 (58.1%)
• 10-turn mixed loop:
  • Raw: 25,520
  • Facade: 10,669
  • Savings: 14,851 (58.2%)

Interpretation: savings are not a narrow edge case. They range from strong (~58%) to extreme (~96%), depending on upstream schema density.

Why Plain MCP Connectivity Loses at Scale

Direct connectivity is great for compatibility, but costly for repeated agent loops:

1. Agents repeatedly ingest oversized tool surfaces.
2. Planning context gets crowded with metadata.
3. Every model/provider integration has to cope with raw heterogeneity.

A compact facade avoids all three:

1. Much smaller default context payload.
2. Detail fetched only for selected capabilities.
3. One normalized envelope for calls, errors, and policy behavior.

This Is Not Just a Token Story

Token savings are the headline. Operational gains are the moat:

• deterministic call/error envelopes
• centralized policy, redaction, and audit controls
• stable contract across CLI, HTTP, and MCP-native clients
• backend MCP server churn decoupled from agent prompt stability

In short: you get both lower spend and better systems engineering.

The “up to 96%” claim is grounded in measured scenarios above, not speculation.

What “Next Level” Means

The first era of MCP proved protocol interoperability. The next era is interface efficiency.

Warmplane makes that shift concrete:

• keep MCP on the backend for breadth
• expose a thin, high-signal capability plane to agents
• pay for detail only when it is actually needed

That is how you go from “tool connectivity” to “production-grade tool economics.”