An autonomous SEO agent that runs on a flat Claude subscription (no API bill)

Checking the source for concrete facts so the paragraphs stay accurate and don’t invent details.DigiHold published seo-agent-pack on GitHub as the working engine behind an autonomous SEO agent that drafts articles and pages on a normal Claude subscription rather than metered API usage. Author Nicolas Lecocq frames the repo as cleaned production files after two months of iteration, not a demo scaffold. Over that period the system drafted 70 articles and 65 pages across 4 sites, with every piece held as a draft until a human publishes. The pack is MIT licensed. On Hacker News the thread sits at 1 point with 0 comments so far.

Each night a scheduled Claude Code run on a Mac picks one keyword from real search data, reads the live top of the SERP, writes under a strict anti-slop rulebook, generates images, and saves a draft. Keywords come only from DataForSEO and Google Search Console; zero verified search volume kills the topic. Before writing, the agent maps the current top 3 pages, lists gaps, and must state in one sentence why its version deserves to beat them, or it drops the topic. Every draft gets a banned-pattern pass plus a humanizer pass drawn from Wikipedia’s “Signs of AI writing” guidance, and every price, date, and number is checked against a live source or cut. Sundays the agent re-reads Search Console and rewrites underperformers. Writing runs on the flat Claude subscription with no per-token API bill; images use a Gemini API key; morning status goes to Telegram. The engine lives in skills/seo-engine/ with five keyword gates, competitor analysis, quick-win hunting, writing rules, a 500-word image-prompt method, and the nightly protocol. Setup is one master prompt pasted into Claude Code that installs helpers, interviews for site config, and wires four secrets from a local env file. Claude Pro is stated as enough for one site.

For engineers and builders this is a concrete loop: scheduled agent, gated keyword intake, SERP-aware drafting, fact checks, draft-only publish, and permanent corrections written back into the rule files. The cost model matters: writing cost is capped at a Claude subscription instead of unbounded token spend, while DataForSEO and Gemini remain the paid external pieces. Human approval is the hard stop, so the system is an assisted production pipeline, not unattended auto-publish. Voice still needs tuning early; new sites without Search Console history get fewer quick wins; quality falls if the agent is pointed at unrelated topic clusters.

In market terms this sits against both generic “AI content” SaaS and raw Claude or Codex chat workflows that lack SERP gates, Search Console feedback, and a nightly schedule. It also competes with API-first agent stacks that bill per token for every draft. By packaging method files plus a setup prompt that builds the system on the user’s machine, DigiHold is selling process and constraints more than a hosted product. That is closer to open-sourcing an operations playbook than launching another content mill.

Practical next steps: clone the repo, run the master setup under Claude Code with DataForSEO, Search Console, Gemini, and Telegram keys, then judge one test draft against your voice and SERP before enabling the schedule. Watch how well the five keyword gates and top-3 competitor pass hold quality after a few weeks of permanent rule edits, whether Telegram draft review stays sustainable at one piece per morning, and how the Sunday Search Console rewrite cycle performs once you have real underperformers to fix. Keep draft-only publish, hard timeouts, and env isolation as defaults.