One twelfth of the tokens via CLI

Request: „Fetch the 40 cards from the board — via the CLI, not the MCP server.“

1. The AI calls a self-built command-line tool (CLI) that connects the ticket system, code hosting, document store and wiki
2. The CLI returns payload only — no tool descriptions, no page wrapper, no metadata
3. The AI continues with the same information, at a fraction of the context cost
4. Rule for all agents: CLI first, then the API, then the browser — if a command is missing, it is added to the CLI rather than bypassed

Queries to a ticket system usually go through an MCP server that drags along tool descriptions, page content and metadata — much of the context window is taken before the actual work starts.

Which sources get connected and who has access is the team's call; the CLI itself is deterministic code, not an agent.

The measurement only became possible once CLI access on Windows had been repaired — login, token storage, certificate chain.

Any organisation wiring AI agents to internal systems. Tokens are money and attention — the obvious optimisation is to feed the model payload only.

MCP servers are convenient but expensive: every request carries tool descriptions, page content and metadata into the context. So we built our own command-line tool that connects the ticket system, code hosting, document store and wiki and returns payload only. The measurement: the same board query with 40 cards, once via MCP, once via the CLI — around 17,000 versus around 1,400 input tokens, roughly one twelfth. That means lower cost per query and, above all, more room in the context for the actual task. For a single ticket, the lean and the expensive route differed between around 3,000 and around 34,000 characters; a full pull of all roughly 70 active risks would simply not have been feasible the expensive way — the access route decides cost and feasibility. It becomes decisive once many workstations share the same brain. The magnitude: a factor of 12, not 20 or 30 percent.