CLI reference

saylent <command> [options]. Every command that spends money prints an estimate and asks to confirm before it spends (skip the prompt with --yes). Run any command with --help for this same reference in the terminal.

saylent audit <domain>

Runs a full audit and writes run.json, report.html, and report.md.

Exit codes: 0 every engine answered, 1 the run failed outright or no engine has a usable key, 2 the run completed but at least one requested engine returned zero successful answers (the report says which, and scores from the remaining engines).

saylent verify <run.json>

Re-asks the exact same frozen question set from a baseline run, against the same engines, and writes a movement report instead of a fresh one.

Writes a new run.json (kind: "verify") and movement.html into a sibling dated directory next to the baseline's own, so re-verifying never overwrites what it compares against. Exit codes match audit: 0 / 1 / 2.

Verify refuses to run when the question templates or the resolved sample count have changed since the baseline was frozen, rather than silently comparing two different question sets. Verify and history is the walkthrough: what it reuses, what makes it refuse, what the movement report says, and how to keep re-checking on a schedule.

saylent gate-check <domain> (alias saylent check)

The command that runs the site check. No keys, no LLM calls, $0. It crawls a handful of pages and reports what AI bots can actually reach:

Saylent · gate-check · example.com

  robots.txt   training: GPTBot allowed · ClaudeBot allowed
               search:   OAI-SearchBot allowed · PerplexityBot allowed
               user:     ChatGPT-User allowed · Claude-User allowed
  live probe   GPTBot 200 · ClaudeBot 403 ⚠ (robots allows -> CDN/WAF override)
  JSON-LD      Organization ✓ · Product ✗
  meta         noai: absent ✓ · noindex: absent ✓
  Result       WARN · 1 blocked-at-CDN · 1 missing schema      (4.8s, $0)

The three bot classes: training crawlers (GPTBot, ClaudeBot) build model weights, not citations - blocking them is a legitimate choice and only warns. Search-index crawlers (OAI-SearchBot, PerplexityBot, and friends) and user-fetch agents (ChatGPT-User, Claude-User) feed live answers directly - blocking either one fails the check. A page allowed by robots.txt that still returns 401/403/429 to the live probe means a CDN or WAF is overriding robots.txt, and the line says so.

Exit code: 1 if anything failed, 0 on pass or warn only - wire it into CI to fail a deploy that regresses AI-bot access.

saylent report <run.json>

Re-renders report.html / report.md from an existing bundle. No network, no LLM call, $0. Useful after hand-editing a bundle, or re-rendering with a newer @saylent/report version.

saylent history <dir>

Lists past runs found in <dir>/<YYYY-MM-DD>/run.json:

  2026-09-09  smoke  present 2/6  recommended 0   $0.51
  2026-09-16  verify present 5/6  recommended 2   $0.44   ▲

saylent keys

Keys resolve in order - environment variable → .env in the current folder → ~/.saylent/config.json - and a later source never overrides an earlier one that already has a value. Keys are never printed, never written into run.json or any report, and every error message the CLI prints has known key values scrubbed out first. There's no --openai-key flag on purpose: a key on the command line ends up in your shell history.

saylent questions <domain>

Prints the buyer questions an audit would ask, and writes them to a file you can edit by hand before spending anything on answers.

Cost: one brand-model call (about $0.02 of your credits, plus a free crawl) unless you pass --no-brand-model, which is $0 and needs no keys at all.

Questions and models is the walkthrough: the shape of the file it writes, how to add, remove and retag a row, what a per-row samples value does, and the three file shapes --questions accepts (a previous run.json, a questions.json, or one question per line in a text file).

Only category and problem rows are scored and count toward the recommended band. Every other type - custom included - is asked, judged and reported in full, and deliberately left out of the band, so nothing you add by hand can inflate your own score.

saylent models

Read-only, $0, no network: prints the model registry exactly as it resolves for the keys, config, and environment on this machine right now - which model serves which role, and whether that choice came from a flag, an env var, saylent.config models, or the shipped default. Also shows which model would judge each engine's answers, and warns if a judge would be the same model that answered (self-judging, not an independent check).

The judge is set with --judge (plus --judge-family when the model name does not say which provider it belongs to); --model <role>=<id> takes one of six roles - brand, drafter, chatgpt, claude, gemini, perplexity. There is no judge role on --model. What each role does, and the examples: Questions and models.

Precedence: --judge / --model role=id flag → MODEL_* env var → saylent.config models → the shipped registry default.

saylent mcp

Serves the same pipeline over the Model Context Protocol on stdio, so an agent can run an audit itself. No flags: everything is set by the environment it is started with.

audit and verify take the same arguments the flags above take, from one shared schema - so the two surfaces can never drift. Each tool states its own cost range in its description, and the estimate is checked against max_usd and the daily spend cap before anything reaches a provider.

Three things the MCP server refuses that the CLI allows, because an agent - not a human - picks the arguments and the working directory:

Provider keys are read from the environment, a .env, or ~/.saylent/config.json exactly as they are for the CLI, and can never be passed as tool arguments. Full setup - the client config block for Claude Desktop and other MCP clients, and the argument table - is on The MCP server.

The run bundle (run.json)

run.json is the lossless format every command reads and writes - BUNDLE_VERSION: 1. Top-level keys: version, run (id, kind, profile, status, brand, the frozen engine set, the model used per role, judge mode, timings, cost, skip - which stages this run skipped, if any), brand_model, questions (the frozen set), answers (every raw answer text and every sampled draw, not just the final verdict), citations, corpus_pages, domain_checks, fixes, scores, health. Nothing is summarized or dropped - this is also what a downloaded bundle from the full app looks like (GET /api/runs/<id>/export?format=bundle), so a CLI run and an app run produce the same artifact. Compatibility is additive only: a bundle with a different version is refused rather than guessed at.

Configuration

Drop a saylent.config.json (or .js / .mjs / .ts, in that priority order when more than one exists) in the folder you run from. Every field is optional; an absent file means the shipped defaults. Precedence for any overlapping value: CLI flag → environment variable → saylent.config.* → shipped default.

See Configuration reference for the complete file with every field, its default, and one table of where each control lives on the CLI, in the environment, in the config file and in the app.

Environment variables (CLI tier)

Every one of these is a DEFAULT: the matching flag always wins, and a bad value is refused with the exact variable name and what was expected, never silently ignored.