MCP server or CLI: what does an agent tool cost per turn?

Updated

Short answer

An MCP server puts its tool schemas into the context of every turn; a CLI puts in the prompt that teaches the agent its commands. Measured on one machine, GitNexus's MCP tools cost about 19,700 tokens per turn and Pixel's CLI prompt about 4,160, 4.7× less. The protocol is not the lever, the size is: semble's and stacklit's MCP servers cost less than Pixel's prompt.

Two ways to pay before the first answer

An MCP server announces its tools to the agent with a schema per tool: a name, a description and the shape of every argument. The agent carries those schemas in its context on every turn, whether it calls the tool or not. A command-line tool has no schema; the agent needs a prompt instead, the text that tells it which command answers which question, and that prompt is carried on every turn too.

Either way the cost is fixed and paid before anything is asked, so it is worth measuring on its own, apart from what each answer costs.

What the measurement shows

GitNexus serves its graph as MCP tools and adds a block to CLAUDE.md; Pixel is a CLI whose agent prompt pixel install puts into each session. The first costs several times the second. But semble and stacklit, both MCP servers with a handful of small tools, cost less than Pixel’s prompt: the number of tools and the length of what each one says decide the bill, not the protocol that carries them.

So the question to ask of any agent tool is how many tokens it keeps in the context, and what it returns for them.

The figures

GitNexus, MCP tool schemas and its CLAUDE.md block~19,700 tok
Pixel, the agent prompt of its CLI~4,160 tok
semble, MCP~980 tok
stacklit, MCP~420 tok
semble and Pixel together~5,100 tok

Every figure above is on the benchmarks page, with its sample size and its method: Against GitNexus, Where a specialist wins.

Where Pixel does not win

Pixel is the lighter one only against GitNexus: semble costs about 980 tokens per turn and stacklit about 420, both less than Pixel. None of these figures counts what an answer costs once the agent asks. A user can switch off individual MCP tools, which shrinks a server's share. pixel mcp serves the same index over MCP; its schemas are not measured, so this page gives no figure for it.