Why CLIs beat MCP servers for AI agents

An MCP server registers all its tools upfront. Connect an API with 60 endpoints, and that's 60 tool schemas sitting in your agent's context from the first message — whether it needs them or not. That's thousands of tokens doing nothing but raising your costs and degrading model performance.

A CLI puts zero definitions in context until the agent asks. It runs --help, finds the subcommand it needs, and drills in. Discovery is on-demand, not upfront. The agent only loads what it actually uses.

Some MCP servers are starting to implement dynamic tool discovery, which closes this gap. But most don't, and the protocol doesn't enforce it. A CLI gives you lazy discovery by default, with no extra engineering.

When an agent has 100 MCP tools available, it has to pick the right one every turn. That's a classification problem that gets measurably harder as tool count grows. Agents pick wrong tools, hallucinate tool names, and make confused multi-tool calls.

With a CLI, the agent has one tool: execute a shell command. The routing logic lives in the CLI's subcommand tree, not in the model's head. The agent navigates with --help the same way a developer would. Models are very good at this.

cli users list --output json | jq '.[].id' | xargs -I{} cli projects list --owner {}

Shell pipes, redirection, loops, scripting — agents that can write shell commands get all of this for free. MCP has no equivalent. Every multi-step workflow has to be orchestrated in your agent code rather than expressed as a one-liner the agent can write itself.

When an agent's MCP tool call fails, you dig through framework logs, reconstruct the payload, and try to figure out what went wrong.

When a CLI command fails, you see the exact command the agent ran, paste it in your terminal, and reproduce the issue in seconds. The entire interaction is visible as plain text.

The main argument for MCP is typed, structured input and output. But a CLI that defaults to JSON output and accepts structured flags already gives you that. The agent sends:

cli users create --email alice@example.com --role admin

And gets back a JSON object. No parsing ambiguity, no unstructured text to deal with.