Data-Driven Benchmark

MCP Token Bloat Benchmark

We audited 11 MCP servers. Here's how much context they waste.

March 2026 · 137 tools analyzed

0
Total Tokens
0
Tools Audited
0
Issues Found
0
Avg Tokens / Tool

Server Leaderboard

Sorted by token cost descending. The GitHub MCP server consumes 74.4% of all tokens across 11 servers.

Server Tools Tokens % Total Issues Relative Size
#1 GitHub 80 20,444 74.4% 50
#2 Filesystem 14 1,841 6.7% 31
#3 Sequential Thinking 1 976 3.6% 2
#4 Memory 9 975 3.6% 9
#5 Git 12 897 3.3% 12
#6 Slack 8 815 3.0% 10
#7 Puppeteer 7 642 2.3% 10
#8 Brave Search 2 374 1.4% 4
#9 Fetch 1 249 0.9% 2
#10 Time 2 215 0.8% 1
#11 Postgres 1 34 0.1% 1
Top 5 Costliest Individual Tools (all from GitHub MCP)
assign_copilot_to_issue 810 tokens
actions_list 714 tokens
projects_write 704 tokens
request_copilot_review 646 tokens
merge_pull_request 610 tokens

Optimization Issues by Rule

The agent-friend optimizer checks 7 heuristic rules. Here's what we found across all 137 tools.

long_param_description
Parameter descriptions over 100 chars
49
verbose_prefix
Redundant phrasing like "This tool..."
29
missing_description
Tool or param with no description
23
long_description
Tool descriptions over 200 chars
20
deep_nesting
Schema depth exceeds 3 levels
8
redundant_param_description
Param description just repeats its name
3
duplicate_param_description
Copy-pasted param descriptions
0

Context Window Impact

What 27,462 tokens looks like as a percentage of popular context windows.

GPT-4o
128K context
21%
Claude 3.5 Sonnet
200K context
14%
Gemini 2.0
1M context
3%
This is JUST tool definitions. Before any conversation history, documents, system prompts, or chain-of-thought reasoning. Every message you send pays this tax again.

Audit Your Own Tools

agent-friend audit measures your tool schemas at build time. agent-friend optimize applies the same 7 rules to shrink them automatically. The only build-time linter for AI tool schemas.

Token Calculator Schema Converter GitHub Repo

pip install agent-friend · Open source · MIT License

Methodology