Rankrat
Self-hosted MCP for owned-site SEO, search analytics, indexing, monitoring, and safe remediation.
Open source Open in the app JSON README (API)
About
Self-hosted MCP for owned-site SEO, search analytics, indexing, monitoring, and safe remediation.
Details
- Kind
- MCP servers
- Topic
- Marketing & analytics
- Publisher
- psyb0t
- Origin
- official
- Category
- ferramentas
- Transport
- local
- Version
- 0.20.0
- Last push
- 2026-08-22T20:04:25Z
- Repository state
- ativo
- Language
- Python
- License
- WTFPL
- Added
- 2026-08-29 04:01:13
- Updated
- 2026-08-29 04:01:13
- Origin id
io.github.psyb0t/rankrat
README
# rankrat <!-- mcp-name: io.github.psyb0t/rankrat --> [](https://github.com/psyb0t/rankrat/actions/workflows/pipeline.yml) [](https://github.com/psyb0t/rankrat/actions/workflows/pipeline.yml) [](https://github.com/psyb0t/rankrat/releases) [](LICENSE) [](https://hub.docker.com/r/psyb0t/rankrat) Every SEO signal lives in its own walled garden with its own dashboard, its own API, and its own bullshit auth dance: Search Console over here, GA4 over there, Bing Webmaster somewhere else, plus Cloudflare, Clarity, PageSpeed, CrUX, GTM, and Bing's backlink data. Point an agent at your rankings and it drowns juggling eight consoles. Rankrat drags all of them behind one self-hosted service, so you tell an agent the outcome you want and it inspects, creates, updates, or rips out provider resources through a single surface — REST and MCP, no dashboards. It's a rat, not a burglar. You hand it the provider accounts; those credentials are the only authority it has. The resource lists in `boundaries.json` are inventory and URL-containment data — not a second permission system pretending to keep you safe. Rankrat writes by default; set `RANKRAT_READ_ONLY=true` and every mutating route and tool disappears from discovery. That's the one switch. It speaks MCP over stdio, MCP over Streamable HTTP at `/mcp`, and a FastAPI JSON API under `/v1/`. Wrapper-managed HTTP gets a bearer; a hand-rolled loopback-only launch can drop it if you know what you're doing. **Status:** alpha. The API and tool surface can still move before 1.0 — minor releases may break things on purpose (documented), so pin an exact release if you need it to hold still. ## Contents - [Capabilities](#capabilities) - [Quick start](#quick-start) - [Run it](#run-it) - [Safety model](#safety-model) - [Documentation](#documentation) - [Agent integrations](#agent-integrations) - [Development](#development) - [Project information](#project-information) ## Capabilities | Area | What Rankrat exposes | | --- | --- | | Google | Search Console analytics, inspection and sitemaps; GA4 inventory, historical and realtime reports, ecommerce, funnels, and audiences; Google Tag Manager containers, workspaces, tags, triggers, variables, versions, and publication; property, sitemap, indexing, ownership, onboarding, and rename writes | | Bing | Search, crawl, indexing, sitemap, backlink, quota, keyword, opportunity, cannibalization, site/submission, and safe content-submission operations | | Performance | PageSpeed, Core Web Vitals, CrUX history, Microsoft Clarity insights, Cloudflare analytics, and isolated local Lighthouse audits | | Site intelligence | Whole-site audits, schema eligibility, internal-link graphs, orphan pages, content opportunities, and cross-provider comparisons | | Ownership and onboarding | Google/Bing checks, provider-neutral DNS verification through Cloudflare, and idempotent GA4/Search Console/Bing onboarding | | Backlinks | Bing Webmaster backlink intelligence for configured sites | | Monitoring and remediation | Persistent monitors and issue history, sitemap/URL resubmission, IndexNow, exact Cloudflare purges, finite cache templates, and managed edge redirects | What the runtime advertises is what it does — no hidden endpoints. A read-only deployment drops every write from discovery; a writable one exposes onboarding alongside the rest. The whole list is the [MCP tool catalog](docs/tool-reference.md) and the generated [OpenAPI document](openapi.json). ## Quick start Install the `rankrat` command, create the credentials for the providers you want, then run it. Docker is the only runtime requirement. ### Install Download the installer, read it, then run it — per-user (no root) or system-wide: ```sh curl -fsSL https://raw.githubusercontent.com/psyb0t/rankrat/main/install.sh -o rankrat-install.sh less rankrat-install.sh bash rankrat-install.sh # per-user -> ~/.local/bin/rankrat sudo bash rankrat-install.sh --system # system-wide -> /usr/local/bin/rankrat ``` The mode auto-detects from who runs it; `--user` / `--system` force it and `--uninstall` removes the command. This puts the `rankrat` wrapper — a readable script that drives the published Docker images — on your PATH. Prefer to work from a source checkout? Clone the repo and use `./rankrat` or the `make` targets instead; see [Getting started](docs/getting-started.md). ### Set up Run setup. It lists every provider, you pick the ones you actually use, and it walks you through creating each credential — the exact console clicks for Google OAuth, the Bing Webmaster key, the Cloudflare token, the Clarity token — then hides every value you paste. No copying token permissions out of a README: ```sh rankrat setup ``` Configuring Google? Paste the one-line Desktop OAuth client JSON at the hidden prompt when setup asks — same as every other credential. Setup validates account access — it does not submit URLs or create properties. Your profile lives in `~/.config/rankrat`; override it per launch with `rankrat --data-dir /absolute/path` or `RANKRAT_DATA_DIR` for a separate profile per account or workspace. Want to read the credential steps ahead of time, or what each provider actually gives you? See [Providers and credentials](docs/providers.md). From a source checkout the setup step is `make setup`. ## Run it ```sh rankrat # MCP over stdio (default) rankrat http -d # REST + Streamable-HTTP MCP + Lighthouse, detached with restarts rankrat upgrade # re-pin to the latest release, pull it, drop the previous image ``` From a source checkout: `make run` (stdio) and `make run-http` (HTTP over Compose, attached). `rankrat setup` pins both `psyb0t/rankrat` and `psyb0t/rankrat-lighthouse` to the latest release tag — recorded, along with the HTTP port and read-only flag, in a `.env` at your profile root — rather than tracking the moving `:latest`. `rankrat upgrade` re-pins the images to the newest release, pulls both, restarts a running HTTP stack (reusing the recorded port and read-only flag), and removes the superseded images; it leaves the port and read-only lines untouched. Pass `--rolling` (or set `RANKRAT_ROLLING=1`) to use the moving `:latest` images for a single run without touching the recorded pin. `RANKRAT_IMAGE`, `RANKRAT_LIGHTHOUSE_IMAGE`, `RANKRAT_HTTP_PORT`, and `RANKRAT_READ_ONLY` still override the recorded values per run. To set the host settings once instead of per run, put them in a permanent host env file at `~/.config/rankrat/.env` (override the path with `RANKRAT_ENV_FILE`). The wrapper reads it before selecting a profile, so it can set `RANKRAT_DATA_DIR` (which profile to use) along with `RANKRAT_READ_ONLY`, `RANKRAT_IMAGE`, `RANKRAT_LIGHTHOUSE_IMAGE`, `RANKRAT_HTTP_PORT`, `RANKRAT_OAUTH_CALLBACK_PORT`, and `RANKRAT_ROLLING`. It only fills a variable that is unset or empty, so a `--data-dir` flag or a real env var still overrides it. This is especially handy for MCP launchers that run `rankrat` with a minimal environment and never see your shell exports. It is separate from the per-profile `.env`: the host file selects the profile and the host defaults, while a profile's `.env` pins that profile's images, port, and read-only flag. Over HTTP, MCP lives at `http://127.0.0.1:8080/mcp` and REST under `/v1/`; stdio needs no port and no bearer. HTTP treats your data directory as its Compose project — it drops a `docker-compose.yml` in there if one's missing and never touches yours if it isn't. Under the hood that's two long-lived services plus one one-shot volume initializer. The wrapper, plain `docker run`, raw Compose, HTTP auth, and client wiring all live in [Transports and deployment](docs/transports.md). Browser audits run in a separate `psyb0t/rankrat-lighthouse` image over a Unix socket, and it never sees a single provider credential. `rankrat` stdio brings up an ephemeral Lighthouse sidecar for the session — same capability as HTTP — and tears it down on exit; the wrapper launches it host-side, so the app stays off the Docker socket and reaches Lighthouse only over the socket. See [Lighthouse](docs/lighthouse.md). ## Safety model - Rankrat is writable by default. Configured account credentials authorize all supported provider operations and all resources those accounts can reach. - `RANKRAT_READ_ONLY=true` removes write REST routes and MCP tools from runtime discovery. It is the only capability switch. - The operator decides whether the caller is a careful human-directed agent or a fully autonomous one. Remote HTTP still uses the configured bearer. - Resource arrays are cached inventory and containment data. Onboarding and discovery update them so retries reuse existing resources. - Secrets are read-only mounts; OAuth records and monitor state have separate writable mounts. See [Configuration](docs/configuration.md) and [Security](docs/security.md) before enabling writes or exposing HTTP beyond loopback. ## Documentation - [Getting started](docs/getting-started.md) — install, file locations, verifying setup - [Configuration](docs/configuration.md) — every boundary field and runtime switch - [Providers and credentials](docs/providers.md) — obtaining each credential and its permissions - [Transports and deployment](docs/transports.md) — stdio MCP, Streamable HTTP MCP, REST, Compose, agent integrations - [Feature workflows](docs/features.md) — operations for common SEO, analytics, backlink, and indexing questions - [Lighthouse](docs/lighthouse.md) — the browser worker and its security limits - [Ownership and onboarding](docs/ownership-and-onboarding.md) — Google/Bing property creation and DNS verification - [Monitoring and remediation](docs/monitoring-and-remediation.md) — persistent monitors and provider-write behavior - [Security](docs/security.md) — trust boundaries and production rules - [MCP tool reference](docs/tool-reference.md) — every discoverable MCP tool - [Troubleshooting](docs/troubleshooting.md) — startup, OAuth, provider, sitemap, and browser-audit failures - [Development](docs/development.md) — build, test, generate contracts, audit dependencies, contribute The REST source is YAML-first: [`openapi.yaml`](src/rankrat/api/openapi.yaml), [`seo-openapi.yaml`](src/rankrat/api/seo-openapi.yaml), and [`free-seo-openapi.yaml`](src/rankrat/api/free-seo-openapi.yaml) generate [`openapi.json`](openapi.json). Agent clients can use the repository's skill, Claude Code and Codex manifests, or OpenClaw integration under [`.agents`](.agents/). ## Agent integrations The [Rankrat skill](.agents/skills/rankrat) works in agents that read `.agents/skills/`. The integrations below cover both MCP stdio and a shared Streamable HTTP server. ### Claude Code ```sh claude plugin marketplace add psyb0t/agents claude plugin install rankrat@psyb0t ``` ### Codex ```sh codex plugin marketplace add psyb0t/agents codex plugin add rankrat@psyb0t ``` Installed through the catalog, the skill invokes as `$rankrat:rankrat`. Codex also discovers it as `$rankrat` directly from a Rankrat checkout. ### OpenClaw ```sh openclaw skills install @psyb0t/rankrat openclaw plugins install clawhub:@psyb0t/rankrat ``` The plugin launches the published image over MCP stdio using the standard `$HOME/.config/rankrat/` layout. Set one absolute `RANKRAT_DATA_DIR` in every client to reuse the same provider accounts, OAuth authorization, inventory, and monitor state from any site repository. To use an already-running shared server, configure OpenClaw for its authenticated `http://127.0.0.1:8080/mcp` Streamable HTTP endpoint. See [Transports](docs/transports.md#agent-integrations). ## Development Everything runs in containers — no "works on my machine" roulette: ```sh make help make lint make test ``` Mocked tests need zero credentials. Live tests are opt-in and pull their one account plus a safe target straight from your profile, so they can't wander off and touch something they shouldn't. Full contributor guide in [Development](docs/development.md). ## Project information [WTFPL license](LICENSE) · [Release history](CHANGELOG.md) · [Dependency attribution](THIRD_PARTY_LICENSES.md) · [Issues](https://github.com/psyb0t/rankrat/issues)