Week 26DAY 4 OF 7

MCP Context Budget

Measure and cap MCP context bloat before agents run

Dan Mercede@danmercede
  • AI
  • Dev Tools
  • Open Source
On the web

About this project

Pitch

MCP Context Budget is a local-first, MIT-licensed CLI for measuring and enforcing MCP tool-surface budgets before coding agents run. MCP servers can quietly consume thousands of context-window tokens through tool definitions alone. Then tool responses add another layer of context pressure. By the time an agent starts working, part of the available context may already be gone. This tool makes that cost visible before runtime. It scans MCP tool definitions, estimates schema-token and response-token cost, selects a lean task-relevant tool set, writes a lockfile, and fails CI when the MCP surface exceeds budget. The goal is simple: treat context hygiene like build infrastructure. Instead of discovering schema bloat mid-session, teams can review it in pull requests, lock the expected budget, and block regressions before they reach agent workflows. Core features: * Scan MCP tool definitions * Estimate schema-token and response-token budgets separately * Select smaller task-relevant tool sets * Write budget lockfiles * Re-check budgets in CI * Export results for review/code scanning * Run locally with no hosted service or proxy The core package is dependency-free and local-first. Nothing leaves your machine unless you explicitly opt into optional local semantic ranking.

Inside the product

Screenshots

Why we're launching this

From the team
I built this because MCP tool surface area is easy to treat as free until the agent session starts getting slower, noisier, or less reliable. The first failure mode I wanted to make visible was schema bloat: too many loaded tools consuming context before the model has done any real work. The second was response bloat: tools with reasonable schemas that still return massive payloads. MCP Context Budget turns that into something reviewable.
Dan Mercede

@danmercede · founder

Discussion

0 comments

Sign in to join the discussion.

No takes yet. Be the first to weigh in.

More in AI

See all →