Free tool · Skill Budget

How many tokens does your agent read before you say hello?

Every skill you install and every MCP server you connect is sent to the model at the start of every conversation. Tick what you run and print the receipt. We measured the real tool schemas and skill text, then converted them to tokens with a ratio calibrated against Claude Code's own /context report.

Start from a setup

MCP servers

Skill packs

Search the 499 most-installed skills on skills.sh

10 skills selected.

How your agent loads things

Count the agent's own baseline. Claude Code's system prompt (2,400) and built-in tools (14,300), from its /context report.Tool search is on. Recent Claude Code versions send only MCP tool names up front and load a tool's schema when it is needed. Many clients still send every schema.

Before hello: 38,173 · Checked bagSee receipt

Cost is only half the question. Does a skill improve the result? Run the free comparison workflow →

What each MCP server costs

Measured 2026-09-25. Tokens for every tool's name, description and input schema, sent at the start of each conversation unless your client defers them.

Server	Tools	Tokens
DataForSEO dataforseo-mcp-server 2.8.10	79	64,129
Notion @notionhq/notion-mcp-server	24	29,396
Chrome DevTools chrome-devtools-mcp	30	9,680
Supabase @supabase/mcp-server-supabase	29	7,023
Playwright @playwright/mcp	25	6,853
Filesystem @modelcontextprotocol/server-filesystem	14	3,144
Git mcp-server-git	12	1,832
Context7 @upstash/context7-mcp	2	1,809
Memory @modelcontextprotocol/server-memory	9	1,637
Sequential Thinking @modelcontextprotocol/server-sequential-thinking	1	1,589
Fetch mcp-server-fetch	1	434
Time mcp-server-time	2	400

Popular skills that load the most text when they fire

From the 100 most-installed skills on skills.sh (2026-09-25). A skill's description is always loaded; its full text only when the agent uses it.

Skill	Installs	Always on	When it fires
design-taste-frontend leonxlnx/taste-skill	519,061	98	29,426
ai-image-generation genmedia-labs/skills	662,792	289	9,611
lark-im open.feishu.cn	727,844	78	8,000
lark-im larksuite/cli	454,587	78	8,000
ai-video-generation genmedia-labs/skills	663,717	327	7,966
hyperframes-cli heygen-com/hyperframes	675,655	195	6,365
lark-mail open.feishu.cn	726,640	74	5,846
lark-slides open.feishu.cn	726,665	94	5,494

How to pick skills without burning tokens

Cut MCP servers first. One server can cost more than a hundred skills.
Keep skills per job, not per star count: about five for the work you do every day.
Remove duplicates. Three skills that all write plans compete for the same moment.
Turn on deferred tool loading if your client supports it (Claude Code's tool search).
Check again after every install. Run /context in Claude Code to see your real numbers.

How we measured

MCP servers: each server was started and its tools read with the MCP tools/list request. Tokens = characters of name + description + input schema ÷ 2.54. DataForSEO 79 tools: 64,129 estimated vs 63.8K in Claude Code /context (within 0.6%). GitHub's MCP server is missing because it needs a real token to list its tools.

Skills: the 499 most-installed of 9,831 skills on skills.sh, via its API. We store only names, sources, install counts and token counts; every skill belongs to its author under its own license. Baseline: Claude Code's system prompt and built-in tools from its /context report. Numbers vary by client and version.

Plain-text version for agents: /skill-budget.md. More: the most popular agent skills, and we set up lean agent stacks for teams.