GitHub cut its default MCP tool count nearly in half. Token cost fell about ten percent. The average tool got seventy-seven percent heavier. Consolidation looked like a diet. It was a rearrangement.
Half the menu, almost the same check
On October 10, engineer ravista published a clean measurement. Same vendor, same protocol, two release binaries, no clever flags. GitHub MCP v0.13.0 (August 2025) exposed 90 tools and burned 12,636 tokens of name + description + input schema on GPT-4o's tokenizer. GitHub MCP v2.0.2 (October 8, 2026 - same code as main that day) shipped 46 tools by default.
Token total: about 11.4k.
Tool count down roughly fifty percent. Token bill down roughly ten percent. Average tool size up about seventy-seven percent.
His one-line diagnosis is the piece worth keeping: Merging tools does not shrink your context. It moves it.
Why the tax shows up before you speak
MCP (Model Context Protocol - the common way agents attach external tools) was designed so the model can see the full menu. A client calls tools/list, gets every tool name, Markdown description, and JSON Schema, then injects that catalogue into the system prompt on every turn.
That is not a one-time setup cost. It is a recurring cover charge.
Community numbers have been converging for months. A Claude Code user filed issue #11364 with seven servers at 67,300 tokens - a third of a 200k window - before the first keystroke. Scott Spence watched his own session warn that MCP tools alone were past 66k, later near 82k. Anthropic's own advanced-tool-use writeups have cited setups where tool definitions ate tens of thousands of tokens, and internal peaks over 134k before optimization.
Stack GitHub with Slack and Sentry and secondary digests put the schema load near 143k of a 200k window. You bought a stadium. The seating chart took most of the seats.
Consolidation is the wrong lever
When a server merges ninety tools into forty-six, the JSON does not evaporate. Nested parameters, enums, examples, and longer descriptions soak up the space the missing names used to occupy. Scott Spence hit the same wall the other way: consolidating search tools cut tool count, but average tokens per tool barely moved until the descriptions themselves got lean.
So the instinct many teams still have - "we will just ship fewer, bigger tools" - is a UI story, not a budget story. The model still pays for the full schema every turn, used or not.
Sam Morrow at GitHub has been blunt about the framing: "MCP eats your context" is a harness problem. His own harness, he says, can run an 86-tool server with near-zero startup context. The protocol can hand you a fat catalogue. The client does not have to swallow it whole.
The fixes that actually move the needle
The 2026 interventions share one idea: progressive disclosure.
- Anthropic Tool Search - load tools on demand when the active catalogue exceeds about 10% of context; reported ~85% reduction while keeping the full library reachable.
- Code Mode / code execution with MCP - present servers as typed SDKs or filesystem-like APIs so the model programs against a thin surface instead of carrying every schema; Cloudflare and Anthropic quotes land in the 98-99.9% range on the definition tax.
- Proxy layers - projects such as Action Hub (announced the same night as ravista's measurement) replace dozens of permanent schemas with a search/load/execute surface and claim ~600 tokens of baseline context in front of 100+ servers.
None of those wins come from counting tools on a marketing page. They come from refusing to pay for the menu until someone orders.
Meanwhile, models sell token thrift
The same day the GitHub MCP numbers circulated, Nous Research dropped a stealth coding and agentic model on Nous Portal named MISSINGNO. - free for a limited time, pitched as highly token-efficient, wired for Hermes Agent. Specs and lab origin stayed dark. The community immediately started blind-testing it.
Two clocks, one week: labs shipping models that brag about thrift, and agent stacks still prepaying tens of thousands of tokens for tool menus the model may never open. The efficiency race is real. Half of it is still stuck in the harness.
What to measure before you "optimize"
If your agent feels dumber after you "connected everything," do not start with another model swap. Count the tokens that exist before the first user message: system prompt, skills list, every MCP tool definition. Heavy setups on X this week reported median first-request loads around 131k once CLAUDE.md files, skills, and MCP schemas stacked.
Then ask the only useful question: of those schemas, how many did the last ten tasks actually call?
If the answer is a handful, consolidation will disappoint you the way GitHub's 90-to-46 cut disappointed the token meter. Lazy load, tool search, code mode, or a thin proxy will not.
Half the tools looked like progress. The bill barely moved. That is the tell.