跳到正文
DEV Community· Nova Ops·· 1 小時前AI 評分55

nova-ops-nim-mcp 評估第三個工具:三個候選方案均未獲採用

Three Candidates for One Tool Slot — Pricing Our MCP Server's Third Tool

AI 導讀

nova-ops-nim-mcp 評估三個第三工具候選方案後,決定維持 v1.0.0 的兩個工具,總 Schema 估算約為每輪 198 tokens。nim_get_job_status 約需每輪 60 tokens,因目前兩個生成工具仍為同步操作而暫緩;nim_upscale_image 約需 76 tokens,因現有工作流尚無需求而延後。

正文

Three Candidates for One Tool Slot — Pricing Our MCP Server's Third Tool

A few days ago I published the audit of our MCP server, nova-ops-nim-mcp: two tools, 790 serialized characters, roughly 198 estimated tokens re-sent on every turn — and the rule we adopted before growing it. Measure every candidate, keep the server's total schema under 1,000 estimated tokens per turn, and keep property descriptions to one short line.

Rules are easy to announce. This post is the rule meeting its first real decision: three candidates have been proposed for the third slot, so I wrote all three schemas, ran the same audit on each (today, Node v24.22.1 — the existing two tools re-measured first and came back identical: 309 and 478 chars, 198 total), and priced them.

Candidate 1: nim_get_job_status — 237 chars, ~60 tokens/turn

One parameter: the job_id a generate call returned. If added, the server totals 1,028 chars / ~257 tokens per turn.

This is the strongest candidate on value: the moment generation becomes asynchronous, an agent without a status tool either blocks, guesses, or re-fires the job. But that conditional is doing real work in the sentence. In v1.0.0 the two generate calls are synchronous — there is no job to poll yet. Paying 60 tokens on every turn for a tool whose precondition doesn't exist is exactly the "just in case" registration the budget exists to stop.

Verdict: first in line, not yet. It earns the slot the day the server goes async, and not a day before.

Candidate 2: nim_upscale_image — 302 chars, ~76 tokens/turn

An image_url, a scale of 2 or 4. If added: 1,093 chars / ~274 per turn. Perfectly affordable — and that's the trap. Cheap is not the same as earned. Nothing in our current workflow is waiting on an upscale; it's a capability we'd register because it exists upstream and might be useful someday.

Verdict: deferred. "Might be useful" is how a 2-tool server becomes a 50-tool server at ~100 tokens a head. When a real job hurts without it, it can re-apply — with the job as its reference.

Candidate 3: nim_create_media — 1,449 chars, ~363 tokens/turn

The fashionable option: one flexible tool with an operation enum (generate image, image-to-video, upscale, restyle) and every parameter for all four. It was supposed to save space by replacing tools. Measured, it does the opposite:

  • It costs more than the entire current server — 1,449 chars against both existing tools' 790 combined.
  • Even if it replaced both tools outright, the server would cost ~363 tokens a turn instead of 198.
  • The reason is visible in the schema: most parameters apply only to some operations, so every description becomes conditional prose — "required for generate_image and restyle, ignored otherwise." The one-short-line rule isn't a style preference; it's a diagnostic. When a parameter needs a paragraph, the tool is doing too much.

Verdict: rejected. One flexible tool turned out to be four tools in a trench coat, charging rent for all four on every turn.

The decision rule, after contact with reality

The budget alone didn't decide any of this — all three candidates fit under 1,000 tokens. What decided it, in order:

  1. A real workflow already hurts without it. Not "could use it" — hurts. (Status: hurts later, when async lands. Upscale: doesn't hurt yet. Kitchen sink: hurts nothing; it is the hurt.)
  2. The measured cost fits with room to spare — because the fourth and fifth candidates are already imaginable, and a budget spent to the ceiling is a budget you'll break.
  3. Every property description stays one short line. The moment you write "only used for…", you're not describing a parameter, you're apologizing for the tool's shape.

So the server stays at two tools and ~198 tokens a turn — not because small is virtuous, but because none of the three candidates earned their rent today. One of them knows exactly what it has to wait for. That's what a budget is for: not saying no, but making "yes" mean something specific.


Method, for the skeptics: estimates at 4 chars/token on the JSON-serialized definitions, same script and method as the first audit; candidate schemas were written for this evaluation and measured today (full run reproduced in our working notes). The server itself is unchanged — v1.0.0, commit 5cd2dd3, still two tools. This post is a decision record, not a release announcement.

來源:DEV Community · dev.to