Use it as a library

Two packages, no Next.js, no Supabase, no database: the same pipeline the CLI and the app both run, and the same renderers that produce report.html.

@saylent/engine - run an audit from your own code

The pure audit pipeline, with zero Next.js, Supabase, Inngest, or React imports - the same runAudit() the CLI and the self-hosted app both call. Supply your own persistence (MemoryDbWriter collects every row in memory; implement the eight-method DbWriter interface for anything else) and your provider keys, and get a full RunResult back with nothing written to disk:

import {
  runAudit, MemoryDbWriter, ADAPTERS, MODELS,
  brandModelCall, drafterCall, judgeCall,
} from "@saylent/engine";

const apiKey = { chatgpt: process.env.OPENAI_API_KEY, claude: process.env.ANTHROPIC_API_KEY };

const result = await runAudit(
  {
    runId: crypto.randomUUID(),
    kind: "audit",
    profile: "smoke",
    brand: { name: "Example Co", domain: "example.com", competitors: ["rival.com"] },
  },
  {
    db: new MemoryDbWriter(),
    ask: (engine, q) => ADAPTERS[engine](q, { model: MODELS[`${engine}Answer`], apiKey: apiKey[engine] }),
    llm: { brandModel: brandModelCall, drafter: drafterCall, judge: judgeCall },
  },
);

console.log(result.scores?.overall); // { answered, recommended, mentioned, rec_rate, mention_rate }

ADAPTERS[gemini|perplexity] need their own key the same way; omit an engine from your ask dispatch and it simply isn't asked.

@saylent/report - render a run without the app

The same composers and React components the app and the CLI's report.html both use - pure functions over plain row arrays, no model call, no network:

import { renderReportHtml } from "@saylent/report/render/html";
import { renderMarkdown } from "@saylent/report/render/markdown";
import { reportDataFromBundle } from "@saylent/report/render/from-bundle";
import { readBundle } from "@saylent/engine/bundle";
import { readFileSync, writeFileSync } from "node:fs";

const bundle = readBundle(readFileSync("run.json", "utf8"));
const data = reportDataFromBundle(bundle);
const html = await renderReportHtml(data, { repoRoot: process.cwd() });
writeFileSync("report.html", html);
writeFileSync("report.md", renderMarkdown(data));

The report package's own composer functions (over plain row arrays, no model call, no network) produce a typed union of block types - text, a stat, a bar chart, a list, a quote, a status line, a page list, a set of moves, key-value pairs - that both renderers above consume. Useful if you're building your own presentation of the data instead of the shipped one.

Both packages are npm workspaces inside this repository (packages/engine, packages/report); import them by subpath (@saylent/engine/judge, @saylent/report/ui/badge) to keep your bundle to what you actually use.