Capafy logo Capafy
Get started
MCP Context Budget Auditor

MCP Context Budget Auditor

by: 7s4r
Audit installed MCP tool-schema token cost; get consolidation plans and task-profile allow-lists so cold starts stop burning context.
#Engenharia#Produtividade#Análise
Avaliação
São necessárias mais avaliações
Vendas
0
Como usar
Executar na Capafy
Também em aplicativos externos
Fornecido pelo publisher
Claude Sonnet 5

MCP Context Budget Auditor

Audit the installed MCP surface on Claude Code, Cursor, or Codex. Estimate tool-schema token cost, score what to keep/merge/defer, and ship task-profile allow-lists plus a remediation plan — before cold starts burn your context window.

Problem

Multiple MCP servers can preload a large tool-schema surface into context before any real work starts. That wastes budget, hurts tool selection, and can hit host caps (e.g. Cursor tool limits).

What you get

  1. Inventory from redacted mcp.json and/or tools/list JSON
  2. Honest schema-token estimates (character / tokenizer approximations — not invented dollar bills)
  3. Scoring — cost, redundancy, overlap, criticality
  4. Task-profile allow-lists — coding-only, research, deploy, docs, …
  5. Remediation plan — consolidate, defer/lazy-load, merge, or disable with your confirmation
  6. Artifacts: BUDGET_REPORT.md, ALLOWLISTS/, REMEDIATION.md

Who it’s for

  • Devs with 4–10+ MCP servers on cold start
  • Indie hackers juggling Cursor + Claude Code + Codex
  • Anyone who needs a structured fix path after seeing MCP schema bloat

How it differs

Not this This skill
Design a new MCP from an API Audits installed surface
Python SDK / prompt-cache spend tools Out of scope
Security / governance auditors Complementary, different job
Free “token budget teaching” prompts Ships allow-lists + remediation
Thin OSS token counters Measurement plus consolidation playbook

How to run

  1. Paste a redacted mcp.json and/or tools/list dumps (strip secrets).
  2. Name host(s) + task profiles.
  3. Receive report + allow-lists + remediation.
  4. Apply locally; optionally re-run for before/after.

Model: Claude Sonnet (latest)
Typical runtime: ~15–35 min (enter 25 on the form)

Limitations

  • Estimates are approximate
  • Host features like tool search / defer_loading may already reduce pain — we account for that
  • We never delete MCP entries without your OK
  • Don’t paste API keys, OAuth tokens, or credentials

Privacy

You supply configs/schemas. Do not upload secrets. Run Online processes what you paste in the session; the skill forbids exfiltrating your configs.


View more