Jev over MCP

Why this pairing is more than a wrapper

An agent's expensive habit is deliberating in tokens. It reasons about whether a claim holds, whether a candidate fits, whether content is safe — each of those is a paragraph of generated thought standing in for a judgement, and the paragraph costs real money and real seconds.

Handing the agent a decision tool changes the economics of its own reasoning. The judgement comes back as a number, in under a second, for a fraction of a cent, with a confidence attached — so the agent can also know when not to trust it. That last part is what an LLM asking itself cannot do honestly: verbalised confidence is not calibrated.

What TypeSafe ships

Worth understanding the distinction: a skill teaches an agent how to use the API it already has access to; an MCP server is the access. If your agent can already make HTTP calls with your key, the skills package may be all you need.

The community servers

The part worth getting right: tool shape

The servers above split into two philosophies, and it is the only decision that really matters if you build your own.

Verb tools win in practice, for the same reason Jev's own option descriptions matter so much: the words are the interface. An agent choosing between screen and rank is making an easy decision; an agent composing a Score rubric from scratch is making a hard one.

Rolling your own

It is a small amount of code, and building it yourself means the tool descriptions match your domain rather than a generic one. The shape, in outline:

// One narrow tool beats one flexible tool. Name it after the decision,
// not after the model — the agent is choosing by the description.
server.tool(
  "screen_candidate",
  "Judge whether a candidate meets the role's hard requirements. " +
    "Returns a probability and a confidence, not a verdict.",
  { candidate: z.string(), requirements: z.string() },
  async ({ candidate, requirements }) => {
    const res = await jev.systemOne({
      model: "jev-latest",
      state: { candidate, requirements },
      questions: {
        meets_bar: {
          type: "noul",
          instructions: "Does the candidate meet every hard requirement listed?",
        },
      },
    });
    const { noul } = res.answers.meets_bar;
    // Hand back the number. Let the agent decide what to do at 0.6.
    return { content: [{ type: "text", text: JSON.stringify({ meets_bar: noul }) }] };
  },
);

Common questions