Track and visualize the context window cost of MCP servers, skill packs, and coding agents.
Your 1M context window isn't unlimited. Before you type a single message, system prompts, built-in tools, MCP servers, and skills have already consumed a portion — and it adds up fast.
context-treemap tracks how much each component costs and visualizes it as a treemap — updated daily, with version-by-version change tracking like a stock ticker.
|
MCP Index
|
Skill Index
|
|
Claude Code — Context Window (1M)
|
Codex — Context Window (1M)
|
MCP Servers (agent-agnostic)
| Server | Tools | Tokens | % of 1M |
|---|---|---|---|
| Playwright | 56 | 42,000 | 4.2% |
| GitHub | 26 | 19,500 | 2.0% |
| Notion | 22 | 16,500 | 1.7% |
| Sentry | 16 | 12,000 | 1.2% |
| Filesystem | 13 | 9,750 | 1.0% |
| Memory | 9 | 6,750 | 0.7% |
| Firecrawl | 8 | 6,000 | 0.6% |
| Slack | 8 | 6,000 | 0.6% |
| Supabase | 7 | 5,250 | 0.5% |
| Linear | 6 | 4,500 | 0.5% |
| Context7 | 2 | 1,500 | 0.2% |
| Figma | 2 | 1,500 | 0.2% |
| Seq. Thinking | 1 | 750 | 0.1% |
| PostgreSQL | 1 | 750 | 0.1% |
| Total | 177 | ~133K | 13.3% |
Skill Packs (Claude Code)
Skills have two cost layers:
- Always-on (description metadata) — loaded into system context just by installing
- On invoke (full SKILL.md body) — loaded only when the skill is called
| Skill Pack | Skills | Always-on | On Invoke |
|---|---|---|---|
| Alireza Claude Skills | 165 | 14,419 | ~379K |
| Everything Claude Code | 118 | 4,639 | ~287K |
| Trail of Bits Security | 60 | 2,470 | ~170K |
| Daymade Claude Skills | 45 | 3,879 | ~112K |
| Jeffallan Claude Skills | 66 | 7,247 | ~82K |
| Anthropic Frontend Design | 17 | 1,690 | ~44K |
| Superpowers Lab | 4 | 196 | ~5K |
Agent System Overhead
| Agent | Model | Context | System Prompt | Built-in Tools | Autocompact | Response Buffer | Total |
|---|---|---|---|---|---|---|---|
| Claude Code | Opus 4.6 | 1M | 3,000 | 16,821 | 33,000 | 8,000 | 60,821 (6.1%) |
| Codex | GPT-5.4 | 1M | 2,500 | 8,000 | — | 100,000 | 110,500 (11.1%) |
- Crawl — GitHub Actions crawls MCP server npm packages daily and extracts tool schemas
- Crawl Skills — Fetches SKILL.md files from GitHub repos and parses description metadata
- Measure — Token counts via Anthropic's count_tokens API (free, accurate)
- Render — D3.js treemap + node-canvas generates PNG images
- Track — Version history is committed to
data/, enabling change detection (▲▼%)
# Install
npm install
# Run everything: crawl MCP + crawl skills + render images
npm run update
# Or run individually
ANTHROPIC_API_KEY=sk-... npm run crawl # MCP servers
GITHUB_TOKEN=ghp-... npm run crawl:skills # Skill packs
npm run render # Generate images- MCP tool schemas — Extracted from npm packages at runtime
- Skill metadata — Parsed from SKILL.md frontmatter on GitHub
- Token counts — Anthropic count_tokens API
- Claude Code internals — Piebald-AI/claude-code-system-prompts
- Community reports — SEP-1576, MCP Tax analysis
- Add an MCP server — Edit
config/servers.jsonand submit a PR - Add a skill pack — Edit
config/skills.jsonand submit a PR - Update agent data — Edit
data/agents/*.jsonwith verified measurements - Report inaccuracies — Open an issue with
/contextcommand output
MIT