Agent commands

Everything on the dashboard runs on DispatchSEO's own servers. This page is about the other half: what the pipeline install actually puts into your site's repo, and how you drive it from Claude Code.

The six slash commands

Each one is a thin shim: it calls the get_instructions MCP tool with a workflow name and follows whatever markdown comes back, word for word. None of them contain the actual playbook.

| Command | What it does | When it runs | | --- | --- | --- | | /seo-setup | Writes .dispatchseo/conventions.md (the site facts file) from your repo's actual stack, theme, and voice, and personalizes the backlink playbook. | Once - right after install, and again any time the site changes enough to re-run it. | | /seo-research | Researches SEO keywords from product knowledge, tracks winners, and queues suggestions. | Weekly, automatically - or run it yourself any time. | | /seo-build | Builds the oldest approved guide suggestion into a PR. Guides only. | Daily, automatically - or on demand. | | /seo-build-tool | Builds the oldest approved tool suggestion into a PR. Tools only. | On approval, or the weekly sweep - or on demand. | | /seo-backlinks | Finds and queues backlink prospects for a keyword or competitor (takes an argument). | On demand. | | /seo-report | Summarizes rankings, GSC trends, and next SEO actions. | On demand. |

The four workflows with no slash command

These fire from a dashboard button or a schedule instead of something you'd type - there's no matching command because nothing about them is meant to be run casually from a terminal.

• install - the one-time pipeline install itself. Fired by the one-line setup.sh command the dashboard shows, or by re-running it from Home's install card. It has to exist before any slash command does, so it can't be one.
• trend-scan - fired by the Scan now button on Home or Trends. Deliberately no schedule - the owner decides when to sweep the niche.
• trend-expand - fired by Get takes on a picked radar subject. The dashboard sends the subject's id and title along with the trigger.
• geo-scan - the weekly AI-visibility scan. Runs Wednesdays, with two same-day retries later if the morning attempt gets dropped or delayed by GitHub's scheduler.

Why the shim design matters

Your repo never holds the actual pipeline logic - only triggers, secrets, and these thin command files. The real playbooks - what "good" looks like, the quality bar, the exact steps for each workflow - live in src/lib/instructions/ on the DispatchSEO backend, and every single run fetches them fresh through get_instructions before doing anything.

That means improving a workflow is one edit on the backend, and every connected project picks it up on its very next run - no commits to a hundred connected repos, no version drift between sites, nothing for you to keep in sync. Each response is stamped with the instructions version it was built under, so if a run ever produces something odd, it can be traced back to exactly which instruction set produced it.

What install actually puts in your repo

• .github/workflows/seo-*.yml - the real GitHub Actions workflows: the daily guide builder, the weekly tool build, auto-merge, tool validation, trend scan, trend expand, the geo scan, weekly research, a token-health check, and a canary that proves the PR machinery works.
• .github/mcp-ci.json - wires up the DispatchSEO MCP (this backend, authenticated with your project's bearer token) alongside DataForSEO's own MCP server, for workflows that need raw keyword and SERP research.
• .github/mcp-validate.json - adds only the Playwright MCP (headless, isolated) for the tool-validation workflow, which needs to click around a real browser rather than call DataForSEO.
• .claude/commands/seo-*.md - the six slash-command shims above.
• .dispatchseo/conventions.md - the site facts file: stack, build command, content directories, design tokens, voice rules. This is what turns generic instructions into "how this specific repo builds."

Two smaller files ride along outside the fetched pack on purpose, so a pack update can never clobber them: .dispatchseo/publish-paths (the allowlist auto-merge checks a guide PR against) and .dispatchseo/pipeline-version (lets a re-install skip rewriting files that are already current).

Just talking to the agent

The slash commands exist for muscle memory and for triggering things from a schedule - they aren't the only way in. The same MCP tools are already loaded, so plain conversation works just as well:

• "Research keywords for me" - the same thing /seo-research does.
• "What should I write next?" - reads the current queue and its order.
• "How are we ranking?" - the same summary /seo-report produces.

Alternate connect path: the Claude Code plugin

.claude-plugin/marketplace.json registers a dispatchseo plugin as another way to connect, alongside the per-project setup.sh script and a manual claude mcp add. It wires the same MCP server through two environment variables - DISPATCHSEO_URL (defaults to https://dispatchseo.com) and DISPATCHSEO_MCP_TOKEN (your project's bearer token) - installed through Claude Code's plugin marketplace flow rather than pasted as a one-line command. Useful if you'd rather manage the connection as a plugin than run the setup script, but it connects to the exact same MCP server either way.