OpenRouter's measurement is striking: for the same task, MCP burns 42 times as many tokens as the command line (CLI). The verdict first: the main reason isn't that the MCP protocol itself is slow—it's that the tool list and the returned results stuff the context full. Checking the list and trimming the returned fields helps more than switching protocols.
Where the 42x Comes From
The first bill is the tool list. An MCP server has to stuff the definitions of all its tools into the context of every request: an unfiltered X platform MCP service with 227 tools takes 68,198 tokens for the list alone; the same service keeping only 3 commonly used tools drops to 2,362 tokens. Seven MCP servers connected at once eat 67,300 tokens before a single question is asked—one third of a 200K context window.
The second bill is the returned results. Reading 20 posts returns a median of 41,250 tokens via MCP; the same data through the CLI, trimmed with jq to 8 fields, comes to a median of 4,069 tokens—90.9% less. The protocol's own overhead is small by comparison: across 190 controlled tests, MCP ran only 1.2x to 2x above the CLI.
Three Steps to Cut Consumption
- Connect only the servers you actually use: for each MCP server, ask whether you'll really use it this week—if not, disconnect it for now; for servers you must keep, use an allowlist to expose only the few tools you need.
- Trim results first: have the server return only the fields you need, or filter on the CLI side with jq and similar tools before feeding anything to the model—don't dump the raw JSON into context.
- Write recurring workflows as skills: put fixed procedures in a skill document for the model to follow, instead of discovering tools and failing by trial and error every time.
There Are Times When MCP Is Cheaper
For tasks like code navigation, which require tracing call chains step by step, MCP measured 2.5x cheaper than the CLI (median 127K vs 316K tokens), because a single call returns structured, linked results and saves repeated searching. So don't abandon MCP wholesale: reading data and checking status is cheaper via CLI, while linked multi-step operations can stay on MCP. One more caution: token counts vary by tokenizer—the figures above use tiktoken's o200k_base, while Claude and Grok count differently. Watch the relative multiples, not the absolute values. For what MCP actually is, see What is MCP? How does it differ from API and plugin calls?; for the source of the 42x figure, see OpenRouter measurement: the same task burns 42x more tokens with MCP than with the CLI.
