Open source finally came for three paid tools — and sent a different bill.

An open-source SEO platform (MIT, 3k stars, v0.0.21) you run yourself: keyword research, rank tracking, competitor and backlink analysis, site audits, plus an MCP server so your agent can query it. docker compose up -d after dropping an API key in .env, or one-click to Cloudflare. It bills itself as "all-in-one SEO for you and your AI agent," and the agent hook is the actually-new part — most SEO tools are dashboards a human stares at, not endpoints a coding agent calls.

Reach for it if you're paying $100-plus a month for Semrush or Ahrefs and use a tenth of it, or want SEO data your agent can pull programmatically. Delete the subscription — but not the spend: the app is free, the data isn't. It needs a DataForSEO API key (pay-as-you-go), so you're trading a fixed monthly seat for metered usage that's cheaper only if you query less than the incumbent assumes. Also v0.0.21, Docker is single-user with no auth by default, and Google Search Console is the optional half that makes it useful.

A local-first markdown editor and LLM wiki (GPL-3.0, 1k stars, v0.19.2) from Inkeep: your notes are plain markdown files on disk, synced via git/GitHub instead of a proprietary cloud, with built-in MCP, skills, and agentic search so Claude, Codex, or Cursor can read and edit the knowledge base directly. npm install -g @inkeep/open-knowledge then ok start, or grab the macOS DMG. The pitch isn't "prettier Notion" — it's a notes store that lives in your repo and that your agent treats as a first-class tool over MCP.

Use it when you want a team wiki or spec store the AI can actually work in, without your knowledge living in someone else's database behind an export button. Delete the Notion/Obsidian-sync subscription; your sync is now git, which you already run. The bill that moves: the "AI-native" features lean on Claude/Codex/Cursor, so you bring your own model keys or desktop apps, and the CLI wants Node 24+. GPL-3.0 also matters if you were planning to fork it into something closed.

A local-first document parser (Apache-2.0, v0.1.1) that extracts structured JSON from PDFs, scans, and images using a model running on your own hardware — NuExtract3 via vLLM by default, no cloud fallback by design. It does semantic extraction (invoice numbers, dates, totals against a JSON schema), not just OCR, and ships three front ends: a Web UI, a REST API, and a CLI. git clone, uv tool install --editable ., parsehawk start. It's a day old, which is exactly why it's here — it's the cleanest statement this week of "the sensitive document never leaves the building."

Reach for it when you're piping contracts, invoices, or medical PDFs through AWS Textract or a Nanonets-style API and the compliance team would rather that data not transit a vendor. Delete the cloud extraction bill; the bill that replaces it is hardware — you need a 16GB+ GPU (Apple Silicon or NVIDIA), it's macOS/Linux only, and at v0.1.1 with breaking changes promised before 1.0, treat it as a bet on the trajectory, not a dependency you pin in production yet.

Off the thread but worth knowing: Nub (github.com/nubjs/nub, MIT, 2.2k stars) is a fast all-in-one Node.js toolkit in Rust — a Bun-style consolidation of the JS toolchain rather than a SaaS replacement, which is why it sits outside today's theme for the second edition running. promptctl (github.com/naya-ai/promptctl) is "git for your AI prompts" — local version/diff/rollback from the CLI, on-theme for owning your own data but a day old and too thin to anchor a pick yet.