AltoRank

# AltoRank: open-source, self-hosted AI SEO content writer (Outrank alternative)

**An open-source alternative to Outrank and Jasper.** AltoRank researches a
keyword, writes the article, scores it and fact-checks it: a figure that cites
a page is looked for on that page, and one with no source is flagged for the
reviewer. Every draft can be copied as Markdown (front matter included) or
HTML, or downloaded as a `.md` file, for you to paste wherever you publish.
Approve it, paste it, then **Mark as published** records the live URL, and the
article is tracked from there: indexing and rank.

AltoRank never publishes to your site without an approval recorded under a
named person. Approve each draft by hand, or set a rule that approves drafts
after a hold unless you hold them (the rule publishes through a CMS connection,
which is built but not yet verified end to end; see [Status](#status)).
Agents can research and draft through the MCP server and the agent API.
Neither can approve, and neither can publish anything a person has not already
approved (the one publish-side call retries a failed publish of an approved
article).

Self-host it free under AGPL-3.0, or use the hosted app at `app.altorank.co`.
Open source, the whole product. No feature-gated tier, no `ee/` directory.

---

## Status

The hosted app at `app.altorank.co` is this repository, deployed from `main`
([docs/deploy.md](docs/deploy.md)). A self-hosted install runs the same code
against your own Supabase and your own API keys. Research, drafting, scoring,
fact-checking, audits, rank tracking and Search Console / Bing sync run today;
the table below has the detail. What is not finished yet:

- **No paying customers, no case studies.** Nothing here has a growth figure
  attached to it, and that is on purpose.
- **CMS publishing is built but not yet verified end to end.** The thirteen
  adapters are in `apps/web/lib/cms/` and the connection form covers all of
  them, but none has been watched publishing to a live site yet, so the
  platform tiles on `/connect` ask you to request an integration instead of
  offering Connect. Every draft can be copied as Markdown or HTML today.
  Details under [What works today](#what-works-today).
- **AI visibility tracking is built but not switched on.** The probe, the
  weekly cron and the `/geo` page exist, but nothing in the app turns tracking
  on for a workspace or adds the prompts it measures, so the sidebar marks it
  "soon" (`apps/web/lib/constants.ts`).
- **The CLI is publishable, not yet on npm.** `packages/altorank-cli` bundles
  `apps/web/scripts/cli.ts` into an `altorank` binary (`npx altorank auth
  whoami` once it is published); until then, `npm run cli` from `apps/web`
  drives the same agent API from a checkout. `skills/altorank/SKILL.md` is the
  skill file a coding agent reads, and it does install with `npx`; see below.

Whether you want to read how it works, run it yourself or send a patch,
everything is here.

## The publishing decision

One gate, two ways through it. `lib/publishing/core.ts` refuses any article
that is not `approved` or `scheduled` with `approved_by` set. Only the server
writes those columns: a signed-in user's own database client cannot make an
article approved, scheduled or live (migration 109), and an edit to approved
text sends it back to review. What can write an approval:

- **A person**, from the editor (`approveArticle`), recorded as `approved_by`.
  The button checks for an active plan and refuses a bare, unsourced figure.
- **A rule a named member of the account set** (`auto_approve`, migration 079): the
  publish cron runs stricter checks than the Approve button - active plan,
  somewhere to publish, no unsourced figure, no named reviewer still to sign
  off, no failing audit item, SEO score at or above the floor - after a hold
  window (default 24h) during which the drafted email carries a Hold link.
  `approved_by` is whoever turned the rule on; `approval_kind` says `auto`.
  Held drafts carry the reason on their own row. See
  `lib/publishing/auto-approve.ts`.

`live` is written after an approved publish, by **Mark as published** on an
approved article, and by one thing that is not a publish: the nightly
found-on-site check (`lib/found-on-site/detect.ts`) marks a draft live when
its text is already on the customer's own site. That records something done
outside AltoRank, sends nothing to a CMS, carries no approval, and can be
undone from the editor.

What cannot write an approval: an agent. **No MCP tool or agent API call
approves an article or publishes a draft** - not disabled, absent. The one
publish-side call, `altorank_retry_publish` (`POST
/articles/{id}/retry-publish`), re-runs a failed publish of an article a
person already approved, and is refused for anything else. The MCP tools are
all registered in `apps/web/lib/mcp/server.ts`, and its server instructions
say so to the agent.

Competing tools ship the words "you stay in control" as copy. The difference
is checkable here: every `live` row is an approved article (by click or by
rule) or one found already published on the customer's own site.

## What works today

| | |
|---|---|
| Keyword research + SERP analysis | DataForSEO |
| Domain audit | 9 readiness checks, crawl, PageSpeed |
| Article generation | research → draft → score → fact-check |
| Brand voice | per-workspace voice profiles |
| Publishing | Markdown and HTML export of every draft; **13 adapters** (below), built but not yet verified end to end |
| Locales | **35** (`apps/web/lib/seo/locales.ts`) |
| Rank tracking | scheduled SERP checks |
| Search analytics | Google Search Console; Bing Webmaster Tools (clicks and impressions per day) |
| AI visibility | **not switched on yet**: built (`lib/geo/`, the weekly `/api/cron/geo`, the `/geo` page) to record whether AI answers name you and who they name instead, but nothing in the app turns it on for a workspace yet |

**Publishing destinations** (`apps/web/lib/cms/adapter.ts`) — thirteen adapters
covering **ten CMS platforms**; WordPress is reachable two ways, and `git` and
`webhook` are publishing targets rather than CMSs. Framer, Ghost, git, HubSpot, Magento,
Notion, Shopify, Webflow, webhook, Wix, WooCommerce, WordPress, and the
WordPress plugin — a second, recommended route to WordPress that installs a
plugin and takes a per-site token instead of an application password
(`apps/web/lib/cms/wordpress-plugin.ts`), which is why WordPress appears twice.

None of the thirteen has yet been verified end to end against a live site, so
`CONNECTABLE_CMS` in `apps/web/lib/cms/connectable.ts` is empty on purpose and
every CMS tile on `/connect` offers "Request integration" instead of Connect;
adding an id to that set brings that tile's Connect button back. The
connection form itself is unchanged and still covers all thirteen: the
page's "New connection" button, a draft's publish panel and
`/connect?connect=<id>` all open it. Without a connection, the editor copies
any draft as Markdown (front matter included) or HTML, or downloads it as a
`.md` file.

## Install the plugin

Claude Code. From a terminal:

```bash
claude plugin marketplace add AltoRank/altorank
claude plugin install altorank@altorank
```

Or inside a Claude Code session: `/plugin marketplace add AltoRank/altorank`,
then `/plugin install altorank@altorank`.

`plugins/altorank/` bundles the hosted MCP server (`https://app.altorank.co/api/mcp`,
OAuth, no API key to copy) with a skill written for those tools. The same folder
carries the portable `plugin.json` + `mcp.json` that ChatGPT and Codex read, and
`.agents/plugins/marketplace.json` lists it for them.

## Install the skill

```bash
npx skills add AltoRank/altorank
```

That puts `skills/altorank/SKILL.md` where Claude Code, Cursor, Codex, Copilot,
Windsurf, Gemini CLI, Cline and the other agents the
[skills CLI](https://github.com/vercel-labs/skills) supports will read it. The
skill teaches an agent to drive the agent API (`/api/agent/v1`): preflight the
key, check a site's readiness, list or suggest keywords, generate a draft into
a human's review queue, edit a draft by find-and-replace, move the content
plan. It never approves, deletes or publishes a draft; the API has no such
calls. Its one publish-side call retries a failed publish of an article a
person already approved.

It needs an API key, created at `/settings/api-keys` in the dashboard and
exported as `ALTORANK_API_KEY`. Self-hosted installs also set
`ALTORANK_BASE_URL`. `skills/altorank/commands/` holds five thin slash
commands for the common calls; the skills CLI does not register slash commands,
so copy them into `.claude/commands/` (Claude Code) or your agent's equivalent
if you want them.

## Running it

Requires Node 22+, and a Supabase project (local via Docker, or hosted).
The full deployment story, hosted and self-hosted, is in
[docs/deploy.md](docs/deploy.md); the Cloudflare assessment is in
[docs/deploy-cloudflare.md](docs/deploy-cloudflare.md).

```bash
npm install
cp docker/.env.example apps/web/.env.local   # every variable, with what each one enables
npm run dev
```

Required before a self-hosted install works end to end:

- your Supabase URL, its anon key and its `SUPABASE_SERVICE_ROLE_KEY`
- `NEXT_PUBLIC_APP_URL`, the address people open the app at. Every emailed
  link (signup confirmation, invites, password reset) is built from it, and the
  Docker build bakes it into the image, so set it before `docker compose up
  --build`; compose refuses to build without it.
- `ANTHROPIC_API_KEY`
- `ENCRYPTION_KEY` and `CRON_SECRET` (`openssl rand -hex 32` for each).
  `docker compose` refuses to start without `CRON_SECRET`, and without it no
  scheduled job runs.
- `RESEND_API_KEY` and `RESEND_FROM_EMAIL` (a sender on a domain verified in
  your Resend account). Signup confirms by email, so without them nobody can
  create an account and nobody can reset a password.

Add `DATAFORSEO_LOGIN` / `DATAFORSEO_PASSWORD` for keyword volume and rank
tracking; without them those steps skip rather than fail. Everything else in
the example file is optional and each one disables exactly one feature. The
file says which. Product analytics (the two `NEXT_PUBLIC_POSTHOG_*` variables,
project token and host) are optional too: without them the browser console
says analytics are off and nothing else changes.

Other entry points:

```bash
npm run test     # vitest, unit tier: mocks only, needs no services
npm run test:db  # vitest, db tier: *.db.test.ts against the local Supabase only
npm run smoke    # research → prompt → model → fact check, against real APIs
```

and from `apps/web`:

```bash
npm run mcp                  # MCP server over stdio, for driving the engine from an AI assistant
npm run cli -- --help        # the agent API from a shell; auth with ALTORANK_API_KEY
npm run readiness -- <domain>  # the agent-readiness checks on their own
```

### Hosted MCP endpoint

The same tools are served at `/api/mcp` (Streamable HTTP) by the running app,
so an MCP client adds one URL as a connector instead of running the stdio
server: `https://app.altorank.co/api/mcp` on the hosted install, or your own
origin when you self-host. Authentication is a bearer token, either an API key
from `/settings/api-keys` or the token the built-in OAuth flow issues:

- discovery at `/.well-known/oauth-authorization-server` and
  `/.well-known/oauth-protected-resource/api/mcp`
- dynamic client registration at `/api/oauth/register` (public clients, PKCE
  S256 only)
- consent at `/oauth/authorize`, owner or admin only, write scope opt-in
- token at `/api/oauth/token`, authorization-code grant, no refresh token

The token a connector receives is an `api_keys` row (migration 080), so it is
listed, expires and is revoked on `/settings/api-keys` like any other key.
Claude Code registers it with
`claude mcp add --transport http altorank https://app.altorank.co/api/mcp`;
ChatGPT and Claude.ai take the same URL as a custom connector.

Neither the CLI nor the MCP server can approve an article or delete anything —
not by configuration, but because the agent API has no such call and no `DELETE`
handler at all. Nor can either of them publish: the single endpoint that reaches
a CMS is `POST /articles/{id}/retry-publish`, which re-runs a publish that
**a human already approved** and that then failed. An article in draft or review
is refused, and the refusal tells the caller to hand it to a person
(`apps/web/app/api/agent/v1/articles/[id]/retry-publish/route.ts:13-35`).

### Database

`apps/web/supabase/` carries a `config.toml` and the numbered migrations, which
apply in order. With the [Supabase CLI](https://supabase.com/docs/guides/cli):

```bash
cd apps/web
supabase start      # local Postgres + auth; applies the migrations, prints the URL and keys for .env.local
```

`supabase start` applies every migration to the local database on first
start (`supabase db reset` re-applies them from scratch). Do not run
`supabase db push` for a local stack: it targets the linked remote project.
Against a hosted project, `supabase link --project-ref <ref>` first, then
`supabase db push`. The app will not get far without the migrations: every
query targets tables they create.

`docker/` has a container setup if you would rather not run Node directly.

## Layout

```
apps/web/               the engine and dashboard (Next.js)
  app/(dashboard)/      dashboard routes
  lib/audit/            domain analysis, readiness checks
  lib/seo/              recommendations, scoring, locales, topical profile
  lib/content/          generation (one implementation, shared by route + cron)
  lib/ai/               fact checking
  lib/cms/              13 publishing adapters
  lib/geo/              AI-answer visibility
  scripts/mcp.ts        MCP server
  scripts/cli.ts        CLI over /api/agent/v1
plugins/altorank/       Claude / ChatGPT / Codex plugin: hosted MCP server + skill
skills/altorank/        the skill file a coding agent reads (`npx skills add AltoRank/altorank`)
packages/altorank-cli/  npm package `altorank`: bundles scripts/cli.ts into a bin
docker/                 container setup for self-hosting
tools/agent-readiness/  standalone agent-readiness scanner
```

## Licence

**GNU AGPL-3.0-only.** Full text in [LICENSE](LICENSE).

In practice: self-host it, run it for your own clients, commercially, and you
owe nothing and publish nothing. The one obligation only reaches people
rebuilding the product. If you modify the source and offer that modified version
to others over a network, your changes have to be shared back under the same
licence.

Running an account on it is free and unencumbered. Relaunching a modified copy as
a closed competing service is not.

## Contributing

Issues and pull requests are welcome. Two things that will get a patch rejected
regardless of how good the code is:

- Adding a way to publish that is not attributable to a person - a click, or a
  rule a named member set and can veto. An agent-triggered publish is the
  canonical example.
- Adding a claim the repository cannot support. If `grep` cannot find the
  feature, the README does not get to mention it.

The code conventions a patch is read against, and how this repository relates
to the marketing site, are in [CONTRIBUTING.md](CONTRIBUTING.md).