An autonomous AI journalist behind hexagonal ports — feed it your data signal, get a researched, fact-guarded, edited article out.
Point it at whatever your product knows (a live dataset, an API, an RSS feed) and it runs the full newsroom loop on top: discover a story worth telling, research it across the web, write it section by section against the evidence, then push it through an editor + fact-integrity gate chain before a single word is published. Domain-agnostic by construction: the engine imports nothing from any host app, and everything brand- or domain-specific arrives through typed ports.
Most "AI blog writers" are a prompt in a trench coat. This is a pipeline with editorial machinery, hardened in production:
- Grounded by construction — sections are written from researched, source-tiered material (wire/gov primary sources ranked first, low-authority hosts down-ranked and labeled; re-reported claims chased to their primary source). An extractive digest layer keeps prompts sharp while the raw corpus remains the guards' ground truth.
- Fact-integrity gates — a fact-guard pass strips fabricated people, scenes, quotes, relationships, and unsourced statistics; a fact-check audit rates every claim against the research and can force a weak article to DRAFT instead of publishing it.
- A real editorial desk — a theme statement recast against what research actually found (stale or dead stories get killed, not published), newspaper line-edit and managing-editor passes with explicit length floors, headline candidates judged against a corpus of editor-written exemplars, structure and corroboration gates.
- Deterministic guardrails around every LLM step — repetition budgets, figure-grounding checks, attribution budgets, title-candidate membership, length-ratio guards. Models drift; the gates don't.
- Total provenance — every prompt, response, search, digest, and gate verdict is recorded per run through a pluggable run-context, so any article can be audited after the fact.
npm i ai-journalistNode 20+ (CI tests 20 and 22), ESM, ships as TypeScript source — consume it via
tsx or your own bundler. Dependencies are small and boring: zod, p-limit,
@openrouter/sdk, firecrawl, rss-parser, date-fns, remark.
A minimal adopter brings four things — a Source (your data), an LlmClient, a SearchClient, and a BrandProfile — and the batteries-included preset assembles the rest:
import { runPipeline } from "ai-journalist";
import { createDefaultInternals } from "ai-journalist/presets";
import { createHttpSource } from "ai-journalist/sources";
import { createOpenRouterLlm } from "ai-journalist/clients/openrouter-llm";
import { createFirecrawlSearch } from "ai-journalist/clients/firecrawl-search";
const source = createHttpSource({ signalUrl: "https://my-api/signal" });
const llm = createOpenRouterLlm({}); // OPENROUTER_API_KEY; dynamic model selection
const search = createFirecrawlSearch({ apiUrl: process.env.FIRECRAWL_API_URL });
const brand = {
name: "My Outlet",
publication: "My Outlet (myoutlet.com)",
beat: "your beat",
bylines: ["A. Writer"],
};
await runPipeline({
source,
sink: { publish: async (post) => ({ url: `out/${post.slug}.md`, status: "DRAFT" }) },
config: { llm, search, brand },
internals: createDefaultInternals({ llm, search, brand, source }),
});Two runnable demos ship in examples/: basic.ts (fully
offline — also the CI wiring proof) and live-minimal.ts (real LLM + search,
writes out/<slug>.md, safely prints SKIP without keys).
ports.ts is the entire customization surface — four public
ports, typed and documented in place:
| Port | What you decide |
|---|---|
Source |
Where signal + grounding facts come from (Http/Rss/File ship; or your own) |
Sink |
Where finished posts land — one publish(post) function, no class to subclass |
EngineConfig |
Which LLM, which search backend, your brand identity/voice, ~70 documented knobs |
Linker |
Optional on-site entity links |
"I want to change X" → CUSTOMIZING.md maps every seam.
Search is fully swappable (SearchClient port): reference clients ship for
Firecrawl (cloud or self-hosted) and self-hosted
SearXNG, with no baked-in hosts.
- Purity — an AST guard fails the build on any
process.envread or hardcoded brand/host literal inside the core: everything host-specific must arrive through the ports. - Prompt stability — 300+ byte-lock checks pin the exact text of every LLM prompt, so prompt drift is a failing test, never a production surprise.
- Wiring — the offline end-to-end example runs under vitest on every push.
Semver, with a CHANGELOG.md section per version. Every
release is published to npm and GitHub Releases from the same tag by CI
(npm trusted publishing, tokenless).
MIT. Extracted from a production news pipeline and maintained as a
standalone, host-agnostic engine — contributions that keep it universal are
welcome (see CONTRIBUTING.md).