Which MCP servers use the most context?

Every tool an MCP server exposes is sent to the model as context before you ask anything. 201 listings have a token-cost estimate; this ranking uses 198 of them. The other 3 are not ranked: Atlassian (Jira/Confluence) listed 0 tools without credentials; REA: Reverse Engineer Anything failed its latest published run (the stored estimate is from an earlier capture); SQLite (npx) failed its latest published run (the stored estimate is from an earlier capture). These are the 20 largest tool lists among the 152 that are not deprecated and not held under our promotion rule (see the note above the table), largest first. It's a size list, not a quality ranking: a large tool list can be worth it if you use those tools.

Looking for small ones? Lowest context cost MCP servers orders measured servers smallest first, using the same token_cost estimate. To add up the servers you run: How much context do your MCP servers cost?

Largest measured tool lists

Listings with no passing tool call and under 50 GitHub stars aren't ranked here unless a check against the vendor's own docs confirmed them. They stay in the full directory. 11 new listings are also held until a tool call passes in our test, whatever their star count. This is a size list, not a score ranking: a no-tool-call cap does not drop a listing from this page if it still qualifies under the promotion rule. Not ranked here: 38. Deprecated servers are left out too (8 measured; see deprecated MCP servers). 10 of the 46 left-out servers have a larger tool list than the last row below, so the numbered list is not a complete list of the largest tool lists; they are named, with their status, below the list.

Ordered by token count, largest first; ties by name. Context tokens = name, description and inputSchema of every tool in the tools/list we captured. Share of a 200k window = context tokens ÷ 200,000. Measured = date we captured that tool list (2026-11-11 to 2026-11-11 for these rows). The other listed servers (132 more) are on the lowest context cost list (smallest first).

What these numbers do and don't tell you

• Tokens are counted with tiktoken cl100k_base. Other models use other tokenizers, so treat every count as approximate for the model you use.
• Counts are for the tool list each server returned without credentials in our sandbox. With your own keys, flags or accounts the list can be longer or shorter.
• Some servers change their tool list with flags, credentials or the connected database. We mark those rows ("Tool list changes with setup"); 17 measured servers have this note in our data.
• 98 listings have a pending token-cost figure (no capture). 201 listings have a token-cost estimate; this ranking uses 198 of them. The other 3 are not ranked: Atlassian (Jira/Confluence) listed 0 tools without credentials; REA: Reverse Engineer Anything failed its latest published run (the stored estimate is from an earlier capture); SQLite (npx) failed its latest published run (the stored estimate is from an earlier capture). A pending figure doesn't mean the list is small. Some of them may have very large tool lists.
• This counts tool definitions only (names, descriptions and input schemas), not the tool results a server sends back when you call it.

Tool lists captured 2026-11-11 to 2026-11-11

How we test: methodology. Data: scored.json (token_cost estimate, tools_count, tools_list_env_dependent). Size only: we don't rate quality here.

FAQ

Which MCP server uses the most context?

Of the 198 estimates this ranking uses, Microsoft 365 MCP (Softeria) (@softeria/ms-365-mcp-server): its tool list is about 238,259 tokens for 202 tools (119.1% of a 200k-token window), measured 2026-11-11 without credentials and counted with tiktoken cl100k_base (approximate for other models).

How many MCP servers use more than a 200k context window just for tools?

2 of the 198 estimates this ranking uses have a tool list over 200,000 tokens, and 9 are over 32,768 tokens (a whole 32k window).

What is the lowest-context pick in each category?

A table on this page lists one server per category (31 categories): the lowest measured tools/list among servers that qualify the same way as the largest list (measured, not deprecated, listed under our promotion rule). It is a size list, not a quality ranking.

How is this measured?

We take the tools array from the tools/list response each server returned in our sandbox without credentials, keep each tool's name, description and inputSchema (what MCP clients usually pass to the model), serialise it as compact JSON and count tokens with tiktoken cl100k_base.

Why is a server I know missing?

98 listings have a pending token-cost figure (no capture). 201 listings have a token-cost estimate; this ranking uses 198 of them. The other 3 are not ranked: Atlassian (Jira/Confluence) listed 0 tools without credentials; REA: Reverse Engineer Anything failed its latest published run (the stored estimate is from an earlier capture); SQLite (npx) failed its latest published run (the stored estimate is from an earlier capture). A pending figure doesn't mean the list is small. Servers whose tool list depends on flags or credentials may also list more tools in your setup than in ours.