ContentOS Agent Lite — open-source AI content agent for founders

ContentOS Agent Lite is an open-source AI content agent for founders that writes from a verifiable process, not a prompt. It walks your coding agent through 8 gates: context → sources → brief → draft → checks → uplift → publish. Every draft gets a transparent 0–100 score across 8 axes, locally, in your terminal. It's a free, self-hosted alternative to one-shot AI writers. It runs in Claude Code / Cursor / Codex with zero API keys, and gets sharper with a free research key. It scores English, Russian, and Arabic content.

Honest scope: this is a skeleton of the content process, not a full platform. It won't guarantee search rankings, do multi-source fact-checking, or run a learning loop. What it gives you: text that is grounded, checked, and in your voice — and a folder you keep and reuse.

Why it's different

• Process, not prompt. Eight gates (00–08). Each emits a score (0–100), a status (READY · REVISE · STOP), blockers (P0 · P1), and the next step.
• No draft before the work is done. The agent may not write until business context, research, the source pack, and the brief all pass. A fact without a source is a hard stop.
• Deterministic where it counts. A tiny optional Python layer scores the draft and flags AI-slop the same way every time — quality isn't the model's mood.
• Runs with zero keys. Inside Claude Code / Cursor / Codex the agent is the model.

Quickstart (60 seconds)

git clone https://github.com/humanswith-ai/contentos-agent-lite
cd contentos-agent-lite

# (optional) the deterministic helpers — needs only Python:
python3 -m venv .venv && .venv/bin/pip install -r requirements.txt

# (optional) free research keys — each optional; Serper needs no card:
cp .env.example .env   # fill any of TAVILY_API_KEY / FIRECRAWL_API_KEY / SERPER_API_KEY

Then in Claude Code / Cursor / Codex, open the content-agent/ folder and say:

"Act as my content agent. Follow gates 00→08. My topic is: <your topic>."

Drop the agent into any of your own projects with python -m scripts.cli init --into <dir>. It scaffolds content-agent/ and teaches your project's coding agent — a managed block in CLAUDE.md / AGENTS.md (re-runnable, your own content preserved) — so you can just say "write a post about …" and it follows the gates.

How it works — the 8 gates

File	Step	Hard stop (P0)
00_FOUNDER_CONTENT_AGENT.md	operating contract	writes a draft before 01–04 pass
01_BUSINESS_CONTEXT.md	ICP / offer / proof / voice / claim policy	no ICP, offer, or claim policy
02_RESEARCH_GATE.md	gather + score real sources	everything rests on weak sources
03_SOURCE_PACK_GATE.md	the citable backbone	a fact with no source
04_BRIEF_READINESS_GATE.md	the brief	missing audience / goal / angle / CTA
05_DRAFT_QUALITY_GATE.md	write from sources, score it	claim not in source pack; score < 55
06_EDITORIAL_UPLIFT_GATE.md	improve — no new facts	uplift invented a fact
07_PUBLISH_READINESS_GATE.md	go / no-go	score < 72 or any P0
08_DISTRIBUTION_GATE.md	publish canonical-first, distribute with backlinks	a published copy with no canonical, or a silo copy with no backlink

Optional deterministic helpers:

python -m scripts.cli research --topic "payment stacks for founders"   # → source-pack draft
python -m scripts.cli score    --file content-agent/_work/draft.md     # → 0-100 + blockers
python -m scripts.cli check    --file content-agent/_work/draft.md     # → anti-slop + score
python -m scripts.cli publish  --file dist/post/index.html --canonical https://you.com/post/  # → pre-publish lint
python -m scripts.cli publish  --file dist/post/index.html --canonical https://you.com/post/ \
  --distribute --title "Headline" --summary "One line."                # → canonical-first drafts

Research uses whichever free key you set: Tavily (search), Firecrawl (read a URL), Serper (SERP / People-Also-Ask, free, no card). With no key, the agent does the research itself.

How it's different from one-shot AI writers

	ContentOS Agent Lite	One-shot AI writers (paste a prompt)
How text is produced	an 8-gate process: sources → brief → draft → checks → publish	one prompt → one draft
Grounding	a fact without a source is a hard stop	will confidently make things up
Quality signal	transparent 0–100 score, 8 axes, you can read the code	a black box
Where it runs	your IDE (Claude Code / Cursor / Codex), local	a web app
Cost	free, open-source (MIT), bring-your-own keys	subscription

This is an honest skeleton — it does the process, not a feature-complete writing suite. For the managed pipeline (per-engine AEO, deep fact-check, best-of-N, team scale, hosting), see below.

Free vs the hosted workspace

	ContentOS Agent Lite (this repo, free)	Humanswith.ai ContentOS (hosted)
The gate process	✅	✅
Research	your key, one pass	managed, multi-tier
Quality score	transparent single heuristic	per-engine AEO (ChatGPT / Perplexity / Gemini / AI Overview)
Fact-check	source-presence discipline	KG + NLI + deep-verify
Best-of-N variants	—	LLM-judge tournament
Publish & distribute	canonical lint + canonical-first drafts	guarded rebuild + post-publish re-scan + Workspace Files sync
Languages	EN / RU / AR content scoring	EN / RU / AR native pipelines
Team scale, hosting, learning loop	self-run	managed

Outgrow the skeleton? → See ContentOS for your category.

FAQ

Is it free?

Yes — MIT-licensed and open-source. You bring your own (free-tier) keys if you want the deterministic research/scoring helpers; otherwise it runs with no keys at all.

Do I need an API key?

No. Inside Claude Code / Cursor / Codex the agent is the model. A free Tavily / Firecrawl / Serper key only makes the research and score steps deterministic and repeatable.

How is this different from the hosted Humanswith.ai ContentOS?

Lite is the open-source skeleton of the process. The hosted ContentOS adds per-engine AEO scoring, KG + NLI + deep fact-check, a best-of-N judge tournament, native EN/RU/AR pipelines, team scale, and managed hosting.

Does it guarantee SEO or AI search rankings?

No. It improves how grounded, structured, and citable your text is — it does not promise positions. Anyone claiming a tool guarantees rankings is selling you something.

Does it work for Russian or Arabic content?

Yes. The scorer and anti-slop scanner accept --lang ru (Russian) and --lang ar (Arabic) — the repo docs themselves are English. Arabic detection covers MSA clichés, «حشو» filler, and dialect markers, and reads the Arabic question mark (؟). Deep native EN/RU/AR pipelines are the hosted ContentOS.

What is AEO / GEO?

Answer-Engine / Generative-Engine Optimization: writing so AI search (ChatGPT, Perplexity, Gemini, Google AI Overviews) can cite you — clear answers, named entities, sourced numbers, structure.

Part of the free Humanswith.ai suite

Three open-source agents (MIT, zero-key) that mirror the loop of the hosted Humanswith.ai Workspace:

• Measure — ai-visibility-probe-lite: where do AI engines ignore you?
• Produce + Publish — contentos-agent-lite (this repo): write from a process, not a prompt — incl. gate 08, publish & distribute.
• Optimize — aeo-site-audit-lite: is your page Fetchable, Chosen, and Extractable?
• Design — brand-card-lite: your brand tokens → on-brand social/OG cards.

Contributing & license

Issues and PRs welcome. MIT — LICENSE. Security: SECURITY.md · verify before install: TRUST.md.

Built by

Gregory Shevchenko, founder of Humanswith.ai. Born from the R-Founders "content agent for founders" workshop · MIT · © 2026 Humanswith.ai