Mercury CLI
Documentation

Providers

Jev

JevEval gives the session model a numerical second opinion from TypeSafe's Jev, through TypeSafe or OpenRouter. The model supplies evidence and asks yes/no, choice or score questions. Jev judges the evidence and answers in numbers.

Set up official Jev access

  1. Open /jev.
  2. Enter your own TypeSafe API key.
  3. Set the spending and request limits you want.
  4. Turn the switch on.

Jev is off by default. /jev, the JEV row in /config and the JEV row in the Boot Menu read and write the same switch. The tool joins or leaves the model’s tool list at the next turn boundary.

For OpenRouter access, sign in or store an OpenRouter key through /logins, then use /jev or on. /jev or off turns Jev off. /jev shows which connection is active; TypeSafe and OpenRouter have separate keys, spend meters and caps.

For official access, the TypeSafe key is held in Mercury’s credential store. TYPESAFE_API_KEY, when set, takes precedence over the stored key. Mercury shows whether a key is present and where it comes from, without displaying its value.

Limits and spend

  • Session allowance: Defaults to $20 for each connection and resets with /clear. Mercury counts this spend itself; the provider balance is separate.
  • Pace: Defaults to 100 requests per minute for the main model.
  • Request ceiling: An optional limit on requests per session; off by default.
  • Crewmates: Access is on by default when Jev is enabled, and can be disabled separately. Each crewmate has 200 calls; together they share a pace of 50 requests per minute and the session allowance.

Spend appears in /jev and as a separate row in /usage. When the screen has no usable session figure, it reports the figure as unavailable.

Questions and results

The tool can compare explanations against evidence, judge measured results or check a proposal against recorded decisions. It accepts named facts, paragraphs, a file of evidence or an inline table, and applies the same questions to every item. Requests run in parallel and return one table, with uncertain answers marked unsure. Choice and score answers are unsure below 0.6 confidence. A yes/no answer has no confidence field: its probability of yes is unsure from 0.4 through 0.6; a low probability outside that interval favours no.

A TSV, CSV, Markdown pipe table or JSONL file supplies one evidence item per row, under the header's names. An inline table supplies columns and rows directly. A text file is one item, or can be split on blank lines with the paragraphs format. Files and tables add at most 100 rows per call, and a file is at most 1,048,576 bytes. An oversized input or a ragged row is refused before anything is sent; nothing is trimmed.

Evidence is sent through the selected TypeSafe or OpenRouter connection. Keep the evidence focused: leave out whole source files, transcripts, secrets and unrelated material. A numerical answer is an opinion that informs your decision; it cannot approve a permission or prove that code passes its checks.

Read the status

/jev distinguishes a disabled switch, a missing or invalid key, local spending or request limits, and provider refusals or outages. The tool reports the reason when a request is unavailable. A temporary limit identifies when another attempt is allowed; a lasting refusal stops further Jev calls for that session while ordinary work continues.