MCP or command line: how to connect an AI agent without waste

“Stop using MCP.” The line is making the rounds on LinkedIn, backed by a test: same agent, same question about a client’s cash position in accounting software, and three times fewer tokens consumed through the command line than through the MCP server. A test on a single task is not a law, but it confirms what we have been saying for a while: MCP servers consume a lot.

The right question is not “MCP or no MCP”. It is: for this agent, on this task, which entry point costs the least and breaks the least?

Where the gap comes from

An MCP server is a standard plug that exposes a piece of software’s actions to an AI assistant. A CLI is a program you launch with a command in a terminal, with its options, and that prints a result.

The consumption gap comes down to three mechanisms.

1. The catalogue loaded upfront. Most MCP clients inject the description of every tool on the server into the agent’s context: invoices, suppliers, entries, reconciliations, when all it wants is the cash position. The richer the server, the higher the bill, and it is paid at every exchange.
2. Full responses. An MCP tool often returns the raw API response, with dozens of fields. The agent has to read everything. With a CLI, it filters before reading: it keeps only the balances and the movements that matter.
3. Round trips. When a task chains several calls, every intermediate result passes back through the model. A command can chain those steps itself and return only the end result.

Anthropic documented the effect on one of its examples: a workflow copying a document from one tool to another went from about 150,000 tokens to about 2,000 when the agent wrote code instead of calling the tools one by one (Anthropic engineering post). It is an example chosen by the vendor, not an average, but the order of magnitude shows where the context goes.

We have seen it ourselves: when too many connectors are plugged in at once, hundreds of tools are described to the agent before it starts, and sessions slow down. Unplugging the ones we did not use was enough to win back time.

When the CLI wins

• The agent runs in a terminal: a coding agent, a scheduled task on a server, a pipeline. The CLI is at home there.
• An official CLI exists and is maintained by the vendor. The major development and cloud tools all have one.
• The task is repeated hundreds of times a day: there, a factor of three on tokens becomes a budget line.
• The result must be verifiable. A person reruns exactly the same command and gets the same figure. In accounting or finance, that is non-negotiable.
• Fewer parts that break: a command works or returns a clear error, there is no server to keep connected.

When MCP remains the right choice

• No terminal. In a mainstream chat assistant, the user plugs in a connector in two clicks; they will never install a CLI. For a sales team querying its CRM, MCP is the only realistic route.
• No CLI. Many business applications do not publish one. Writing and maintaining your own costs more than a few tokens.
• Access to keep tight. An MCP server only exposes the actions you chose, with the user’s own login. A CLI in a terminal often gives access to much more, and has to be fenced in: limited-rights accounts, secrets out of the agent’s reach.
• Clients are improving. Recent tools no longer load the whole catalogue: they look up the useful tool and describe only that one. The gap narrows for whoever picks the right client.

How to decide, for a business

1. List your agents’ repeated tasks and measure what they consume. Without measurement, the debate stays theoretical; our article on AI costs and model routing shows how to track it.
2. Unplug the connectors you do not use. The first gain needs no development: every connected server weighs on every exchange.
3. Prefer the CLI for agents that run on their own on a server, when the vendor publishes one.
4. Keep MCP for teams that work in a chat assistant, with rights set as for a person: that is the subject of our guide to talking to your CRM in natural language through MCP.
5. Either way, keep the keys out of the agent’s reach: our article on a secrets vault for AI explains how.

Our view

“Stop using MCP” is a good headline and bad advice. MCP made agents accessible to teams who will never open a terminal, and that is real progress. But it has a hidden cost, paid in tokens at every exchange, and many businesses plug everything in without measuring it. Our rule: for an agent that runs on its own and often, the CLI first when one exists; for a team that talks to its tools, MCP, with few connectors and tight rights. And in every case, measure before deciding.

Sources

1. Anthropic engineering post (new tab)anthropic.com