Your agent burns 100,000 tokens before you type a word
June 17, 2026 · Wu Jun, The Beauty of Mathematics~5 min read
Open a fresh chat with your AI coding agent. Connect it to the tools your team actually uses — GitHub, Jira, Slack, a database. Now, before you have typed a single character, look at the context meter. On a typical enterprise setup it has already spent somewhere between 100,000 and 200,000 tokens. Not on your problem. On reading the instruction manuals for tools it may never touch. This is the quiet scandal underneath the loudest infrastructure fight of 2026 — and it has a shape you can measure.
The receipt nobody reads
When an agent connects to an MCP server — the Model Context Protocol, the standard way to plug tools into a model — every tool definition that server exposes is loaded into the context window. All of them. A single GitHub MCP server with 93 tools costs roughly 55,000 tokens just to say hello; Jira adds another 17,000. One developer's real stack — Linear, Notion, Slack, Postgres — burned 21,000 tokens before any work began, which is 10.5% of Claude's 200K window gone to overhead. And the benchmark that made people gasp: Scalekit asked one trivial question — "what language is this repo?" — and clocked 1,365 tokens through a command-line tool versus 44,026 through MCP. Same answer. Thirty-two times the price.
MCP isn't the villain. Scarcity is.
It would be easy to read all this as "MCP bad, CLI good," and a loud corner of Hacker News did exactly that. But the more careful consensus is subtler, and it isn't about the protocol at all — it's about context. A context window is finite. Every token you spend describing a tool you won't use is a token you can't spend reasoning about the tool you will. MCP buys you real things: authentication, governance, an audit trail, access for non-technical teammates. For an enterprise, that overhead is the price of admission. For a solo developer wiring up a coding agent, it's mostly dead weight. The fight isn't good-versus-evil; it's a budget problem wearing a protocol costume.
Wu Jun: you're not paying in tokens. You're paying in bits.
Decades ago, in The Beauty of Mathematics, Wu Jun walked readers through the one idea that quietly runs all of modern computing: information theory. Claude Shannon's insight was that information can be measured — in bits — and that any channel, from a copper wire to a conversation, has a finite capacity. You cannot push more meaning through it than its capacity allows; try, and the surplus is just noise. Wu Jun's gift is the translation: a communication channel is a pipe, and what flows through it isn't words, it's the reduction of uncertainty. The width of the pipe is fixed. What you choose to send through it is not.
Hold that next to an AI agent and the whole debate snaps into focus. The context window is the channel. Its capacity is fixed — 200,000 tokens, say. Tool definitions are traffic in that pipe. When you preload 150,000 tokens of tool manuals, you haven't made the agent smarter; you've stuffed the channel with low-information cargo and left almost no room for the signal that actually matters — your question, the code, the reasoning. The agent gets dumber not because the model shrank, but because the pipe is full of packing peanuts.
When an agent connects to an MCP server, every tool definition it exposes is loaded into context — all of them, before you ask anything. One real stack hit 21,000 tokens (10.5% of Claude's 200K window); Scalekit benchmarked one repo query at 1,365 tokens via CLI versus 44,026 via MCP — a 32x gap; a 93-tool GitHub server alone costs ~55,000 tokens at startup, and a typical enterprise stack burns 100,000–200,000 before a single character is typed. Wu Jun's The Beauty of Mathematics frames it cleanly: a context window is a finite channel, and the true currency isn't tokens but information per token. Load the signal you need, not the whole inventory. The 2026 consensus: CLI for most coding work; MCP's overhead is the price of enterprise governance.
Information per token, not tokens
Here is the metric Wu Jun's lens hands you, and it's the one the token-counting headlines miss. The currency of an LLM was never tokens. It's information per token — how much uncertainty each token in the window actually removes. By that measure, a 600-token definition for a tool the agent never calls has an information value of roughly zero. It's pure noise occupying premium space. The same instinct drives one of the book's favorite examples, TF-IDF: a word that shows up everywhere carries almost no information; the rare, on-point word carries a lot. Loading all tools, always, is the all-words-are-equal mistake. Loading only the tools this task needs — on demand, the way a command-line agent does — is the search engine's instinct: weight by relevance, and spend attention where the information density is highest.
What this means for what you ship
So you don't have to pick a tribe. You have to budget a channel. Audit what your agent loads before it does anything — if a stack eats six figures of context at hello, that's not a feature, it's a leak. Prefer load-on-demand: tools fetched when the task calls for them, not dumped in at startup. Reserve MCP's overhead for where its governance actually earns its keep, and reach for lean CLI tools for the 90% of coding work that needs no audit trail. And the next time a vendor brags about how many tools their server exposes, remember Wu Jun's quiet rule: a longer menu is not more information. It's usually just more noise in a pipe that was never going to get any wider. Spend the channel on signal.
A bigger toolbox isn't more information.
It's usually just more noise in a pipe that can't get any wider — and the currency of an agent was never tokens, but how much each token actually tells it.
Framework drawn from Wu Jun's The Beauty of Mathematics (吴军《数学之美》) — information theory, channel capacity, and TF-IDF. Token figures from the 2026 MCP-vs-CLI debate: one real stack at 21,000 tokens (10.5% of Claude's 200K window); Scalekit's benchmark of 1,365 (CLI) vs 44,026 (MCP) tokens, a 32× gap; a 93-tool GitHub server at ~55,000 tokens and a typical enterprise stack at 100,000–200,000 on initialization. Popular tech commentary; figures per the original reports.
これを AI エージェントの隣に置くと、論争全体の焦点が一気に合う。文脈窓がその通信路だ。容量は固定——たとえば20万トークン。ツール定義はそのパイプの中の交通量だ。15万トークンのツール説明書を先読みしても、エージェントが賢くなるわけではない。低情報の貨物で通信路を埋め、本当に大事な信号——あなたの問い、コード、推論——にほとんど余地を残さなかっただけだ。エージェントが愚かになるのは、モデルが縮んだからではない。パイプが緩衝材で埋まっているからだ。