The Context Budget Index · The MCP Context Cost Index · November 2026 · updated monthly · counted with ctxcap
The MCP Context Cost Index
Every MCP server a client connects sends its tool definitions in every session, before the model has been asked anything. The definitions are billed to the context window, and most people never see the bill. This page is the bill.
- 16servers measured over stdio
- 281tool definitions between them
- 1,074tokens in the median server
- 46,512in the largest, Sentry
This is not a ranking of quality. A server with 73 tools can do more than one with two, and a large schema is sometimes the honest way to describe a large API. What the table measures is what it costs to have that server switched on: tokens that sit in the context window for the whole session. A server nobody calls still pays.
The comparison column is against ctxcap's own server, which exposes four tools: measure, plan, verify, and trim. It exposes four tools and sits near the bottom of the table at 507 tokens. That is the point of the reference in the README: a server that answers four questions can cost 507 tokens, while one that answers many can cost 46,512 before it is asked anything.
This is the permanent page for November 2026. The newest month is here.
16 servers, largest first
| # | Server and package | Tools | Tokens | × ctxcap |
|---|---|---|---|---|
| 1 | Sentry@sentry/mcp-server@0.49.0 | 73 | 46,512 | 92× |
| 2 | Firecrawlfirecrawl-mcp@3.40.0 | 32 | 17,764 | 35× |
| 3 | Notion@notionhq/notion-mcp-server@2.6.1 | 24 | 17,163 | 34× |
| 4 | Chrome DevToolschrome-devtools-mcp@1.13.0 | 33 | 6,753 | 13× |
| 5 | Playwright@playwright/mcp@0.0.85 | 25 | 3,936 | 7.8× |
| 6 | GitHub@modelcontextprotocol/server-github@2025.4.8 | 26 | 3,548 | 7.0× |
| 7 | Filesystem@modelcontextprotocol/server-filesystem@1.0.0 | 14 | 1,695 | 3.3× |
| 8 | Everything@modelcontextprotocol/server-everything@1.0.0 | 13 | 1,082 | 2.1× |
| 9 | Context7@upstash/context7-mcp@4.7.0 | 2 | 1,065 | 2.1× |
| 10 | Exaexa-mcp-server@3.4.3 | 3 | 1,058 | 2.1× |
| 11 | Memory@modelcontextprotocol/server-memory@1.0.0 | 9 | 957 | 1.9× |
| 12 | Slack@modelcontextprotocol/server-slack@2025.4.25 | 8 | 681 | 1.3× |
| 13 | Puppeteer@modelcontextprotocol/server-puppeteer@2025.5.12 | 7 | 540 | 1.1× |
| 14 | ctxcapthis repository | 4 | 507 | 1.0× |
| 15 | Brave Search@modelcontextprotocol/server-brave-search@0.6.2 | 2 | 319 | 0.6× |
| 16 | Browserbase@browserbasehq/mcp@3.0.0 | 6 | 310 | 0.6× |
Measured 2026-11-16. 310 tokens is the smallest: Browserbase. The largest is 46,512, 92× ctxcap's server. The data.
Not measured
These are popular servers, but there is no honest single number for them. They are listed with the reason rather than dropped.
- Supabase (
@supabase/mcp-server-supabase@0.17.0) — Exposes no tools until a project is linked, so there is nothing to measure at startup. - Cloudflare (
@cloudflare/mcp-server-cloudflare@0.2.0) — Requires a product name as a subcommand; there is no single default tool list. - Redis (
@modelcontextprotocol/server-redis@2025.4.25) — Connects to Redis before answering initialize, so it does not start without a reachable database.
How it was counted
- Each server was installed from npm at the version pinned in
scripts/mcp-servers.json, started over stdio, sentinitializeandtools/list, and stopped. No tool was called and no conversation was started. - The count covers
name,description, andinputSchemafor every tool. It leaves outtitle,annotations, andoutputSchema, which some clients also send, so every number here is a lower bound. - Credentials were placeholders.
tools/listis answered without a real token by these servers; an API call is not needed to learn what the tools cost. - Some servers can hide tools behind configuration or a selected toolset. The number is the default tool list, which is what a session pays before anyone tunes it. Sentry notes that it can be narrowed.
- Tokens are counted with
gpt-tokenizer@3.0.1, the same tokenizer as the The Context Budget Index. Other models tokenize differently, so read these as approximate and compare them with each other. - Servers that connect to a database before answering, expose nothing until configured, or require a product subcommand are in the list above, not silently absent.
What to do about it
The cheapest MCP server is the one that is not connected. The second cheapest is one that exposes the tools a session needs and no more: disable toolsets you do not use, and prefer a server that can return its tool list on demand rather than one that sends all 73 at startup.
npx ctxcap --compare # your AGENTS.md, against the Context Budget Index npx ctxcap init --agents mcp # four tools, about 500 tokens
Data for this month: 2026-11.json. Method and how to rebuild it. Found a server that should be here? Open an issue.