npm package discovery and stats viewer.

Discover Tips

  • General search

    [free text search, go nuts!]

  • Package details

    pkg:[package-name]

  • User packages

    @[username]

Sponsor

Optimize Toolset

I’ve always been into building performant and accessible sites, but lately I’ve been taking it extremely seriously. So much so that I’ve been building a tool to help me optimize and monitor the sites that I build to make sure that I’m making an attempt to offer the best experience to those who visit them. If you’re into performant, accessible and SEO friendly sites, you might like it too! You can check it out at Optimize Toolset.

About

Hi, 👋, I’m Ryan Hefner  and I built this site for me, and you! The goal of this site was to provide an easy way for me to check the stats on my npm packages, both for prioritizing issues and updates, and to give me a little kick in the pants to keep up on stuff.

As I was building it, I realized that I was actually using the tool to build the tool, and figured I might as well put this out there and hopefully others will find it to be a fast and useful way to search and browse npm packages as I have.

If you’re interested in other things I’m working on, follow me on Twitter or check out the open source projects I’ve been publishing on GitHub.

I am also working on a Twitter bot for this site to tweet the most popular, newest, random packages from npm. Please follow that account now and it will start sending out packages soon–ish.

Open Software & Tools

This site wouldn’t be possible without the immense generosity and tireless efforts from the people who make contributions to the world and share their work via open source initiatives. Thank you 🙏

© 2026 – Pkg Stats / Ryan Hefner

context-xray

v0.2.1

Published

See what your agent setup costs you. Weighs the context-window tokens your MCP servers, CLAUDE.md, rules files and skills add to every request - before you type a word - across Claude Desktop, Claude Code, Cursor, Windsurf, VS Code, Zed, Cline, Roo Code,

Vulnerabilities

Powered byPowered by Snyk
  • 0High
  • 0Medium
  • 0Low

Links

Git

Maintainers

beeeeenbeeeeen

Keywords

mcpmodel-context-protocolmcp-servercontext-windowtoken-usagetoken-costtokenscontextclaudeclaude-codeclaude-mdcursorcursorruleswindsurfzedclinegemini-clicodexskillstoken-budgetllmai-agentscostauditclianthropic

Readme

context-xray

See what your agent setup costs you — before you type a word.

CI npm license

npx context-xray

Every MCP server you configure injects its full tool catalog — every name, every description, every JSON schema — into every single request your agent makes. Your CLAUDE.md, your Cursor rules, your skills listing ride along too. You never see any of it. It happens before your first word, it is resent on every message, and it never shows up itemised on a bill.

context-xray finds every server configured in Claude Desktop, Claude Code, Cursor, Windsurf, VS Code, Zed, Cline, Roo Code, Gemini CLI and Codex CLI, connects to each one, weighs exactly what it charges you — and weighs the instruction files those hosts load on their own:

  context-xray  your MCP servers add ~14,443 tokens to every request
  1 config searched, 4 servers found, 4 measured

  rank  server                host               tools    tokens  share
  #1    github                Claude Code           26    ~5,123    35%  ######............
  #2    playwright            Claude Code           24    ~4,785    33%  ######............
  #3    filesystem            Claude Desktop        14    ~2,333    16%  ###...............
  #4    everything            Claude Desktop        16    ~2,202    15%  ###...............

  github  github-mcp-server v0.6.2  26 tools, connected in 3523ms
         513 tok  create_pull_request_review                desc 8, schema 482
         326 tok  list_pull_requests                        desc 10, schema 296
         292 tok  create_pull_request                       desc 11, schema 259
       3,992 tok  … 23 more tools

  playwright  Playwright v1.63.0  24 tools, connected in 3900ms
         429 tok  browser_take_screenshot                   desc 31, schema 376
         324 tok  browser_fill_form                         desc 6, schema 297
       4,032 tok  … 22 more tools

  agent files  instruction files your editors and CLIs load on their own
      ~2,412 tok  CLAUDE.md (project)                 Claude Code
        ~713 tok  .cursor/rules/style.mdc             Cursor
         ~91 tok  skills listing (global)             Claude Code -- 4 skills; bodies load on demand

  ------------------------------------------------------------------------
  ~17,659 tokens on every request (MCP ~14,443 + agent files ~3,216) = 8.8% of a 200,000 context window, before you type a word
  at 200 requests/day and $3.00/MTok input: ~$318/month of uncached input spend

  ! over 10% of the window goes to tool definitions -- disable the servers
    you are not using today

Those are real numbers from real servers — that four-server setup is a perfectly ordinary one, and it spends 7% of the context window on standby.

No install, no config, no account, zero dependencies. It reads the configs you already have.

Why this matters

Context. The context window is the scarcest resource an agent has. Tool definitions are pure overhead: they crowd out your conversation, your files, your actual work. Long sessions degrade sooner, and "compact" happens earlier, in direct proportion to this number. Most people have never seen it.

Money. Input tokens are billed per request. A 14k-token tool catalog at a few hundred requests a day is real spend — caching softens it but does not erase it, and cache writes bill at a premium. The table above is the line item your invoice never shows.

Model quality. Models pick tools by reading their descriptions. A 24-tool server whose every schema looks alike does not just cost tokens — it measurably degrades tool selection. The heaviest tools in the ranking are usually also the ones confusing your agent.

Usage

# find and weigh everything configured on this machine
npx context-xray

# just one host's config, or one server
npx context-xray --config ~/.claude.json
npx context-xray --server github,playwright

# weigh a server that is not configured anywhere yet
npx context-xray -- npx -y @playwright/mcp@latest
npx context-xray --url http://localhost:3000/mcp

| flag | | |---|---| | --precise | exact counts via the free Anthropic count_tokens API (needs ANTHROPIC_API_KEY) | | --json | machine-readable report | | --save <file> | also write the JSON report to a file (a baseline for --diff) | | --diff <file> | compare against a saved report: what got heavier, what changed | | --budget <tokens> | exit 3 when the per-request total exceeds the budget — the CI gate | | --no-static | skip the instruction-file weighing | | --requests-per-day <n> | volume assumption for the cost line (default 200) | | --price <usd> | $/MTok input for the cost line (default 3.00) | | --top <n> | tools listed per server (default 3) | | --timeout <ms> | per-server timeout (default 15000) |

Configs it knows how to read: Claude Desktop, Claude Code (~/.claude.json, per-project entries, and ./.mcp.json), Cursor, Windsurf, VS Code (both the mcpServers and servers shapes, including ${input:...} entries — those are reported as unmeasurable rather than silently skipped), Zed (context_servers, including the nested command object), Cline, Roo Code, Gemini CLI (~/.gemini/settings.json, including httpUrl remotes), and Codex CLI (~/.codex/config.toml — yes, the TOML one). A server configured in several hosts is measured once and attributed to all of them.

The other half of the bill: agent files

MCP servers are not the only thing riding along on every request. The instruction files your hosts load on their own — CLAUDE.md (global and project), CLAUDE.local.md, AGENTS.md, GEMINI.md, .cursorrules, .cursor/rules/*.mdc, .windsurfrules, .github/copilot-instructions.md — are weighed and added to the total. Cursor rules scoped by globs (or left to the agent) are listed but kept out of the every-request total, because that is how they actually load.

Claude Code skills bill in two parts, and the report keeps them honest: every skill's name + description line sits in the listing on every request (counted), while the SKILL.md body only loads when the skill is invoked (shown as "load on demand", not counted).

CI: put a number on it and keep it there

# fail the build when the per-request context bill crosses 8,000 tokens
npx context-xray --budget 8000

# save a baseline on main, diff against it in the PR
npx context-xray --save .xray-baseline.json
npx context-xray --diff .xray-baseline.json

--diff tells you exactly what moved: servers added or removed, tools that appeared, disappeared, or got heavier, instruction files that grew. Exit codes are CI-friendly: 0 all measured, 1 some servers failed, 2 usage error, 3 over budget.

Exact numbers

By default the token counts are estimates from a tokenizer-calibrated character model (JSON weighs heavier than prose, and is marked ~). For exact numbers:

ANTHROPIC_API_KEY=sk-... npx context-xray --precise

This uses the Anthropic count_tokens endpoint, which is free of charge. Only the tool definitions being counted are sent — never your conversation, never your files.

What it does and does not do

  • It connects, performs the MCP handshake, reads the tool/resource/prompt lists, and disconnects. It never invokes a tool.
  • Environment values in your configs are passed to the servers they belong to, and are never printed or transmitted.
  • Instruction files are read locally and only their token weight is reported. File contents never leave the process.
  • Servers that fail to start are reported with the reason (and their stderr), not skipped — a server that cannot start is costing you a different way.
  • The measured tax covers what hosts inject per request: tool definitions plus server instructions. Resources and prompts are listed for information but are not part of the per-request tax.

Programmatic use

import { discover, weighAll } from 'context-xray'

const { specs } = discover()
const weights = await weighAll(specs, {
  timeoutMs: 15000, precise: false, requestsPerDay: 200,
  pricePerMTok: 3, top: 3, concurrency: 4,
})
for (const w of weights) console.log(w.spec.name, w.taxTokens)

See also

The rest of the toolchain, built on the same zero-dependency MCP client:

  • mcp-wtf — your MCP server won't connect; find out why in 10 seconds.
  • mcp-probe — conformance and robustness tests for MCP servers, built to run in CI.

mcp-wtf answers "why won't it connect", context-xray answers "what is it costing me", mcp-probe answers "will it break my users".

Newsletter

Agent Receipts — weekly, first-hand numbers from building and shipping tools like this one: what I measured, what broke, and what I got wrong. Subscribe.

License

MIT