fullvision

FullVision for Claude Code

Revenue-attributed marketing operations in your terminal. FullVision joins ad spend, web behaviour, SEO and Stripe revenue into one attributed dataset; this plugin turns that dataset into changes you can review and apply.

Meta's own MCP can tell you your CPA. Only FullVision can tell you LTV-adjusted ROAS by campaign, attributed to real Stripe revenue, across platforms, on comparable attribution windows. Every skill here exists to use that.

Install

/plugin marketplace add FullVision-Team/fullvision-plugin
/plugin install fullvision@fullvision-plugin

Then, in the repository that serves your website:

/fullvision:fv-onboard

That is the only instruction, and it works from a standing start — no account, no API key, nothing connected. fv-onboard walks you through signup, installs the tracker into your actual codebase as a pull request, connects Stripe and your ad platforms, and does not call a step done until it can see the data arriving.

Already set up? /fullvision:fv-capabilities tells you what works right now with what you have connected, and names the one thing worth connecting next.

Setup

Nothing to do up front — fv-onboard runs the login for you at the right moment. To connect a machine on its own (a second laptop, a re-login after 90 days, switching workspace):

/fullvision:fv-login

This opens your browser, you pick a workspace and click Authorize, and the key is saved to ~/.fullvision/credentials.json. It is read-only, expires in 90 days, and is revocable any time in FullVision → Settings → API keys.

Prefer to manage the key yourself? Export FULLVISION_API_KEY=sk_… instead — it takes precedence over the saved credential.

Optional, per server you want to use: GOOGLE_ADS_DEVELOPER_TOKEN, LINKEDIN_ACCESS_TOKEN, BREVO_API_KEY. Meta and Webflow authenticate in the browser on first use. You do not need all of them — every skill degrades gracefully.

Recommended: enable deferred tool loading (ENABLE_TOOL_SEARCH). This plugin bundles six MCP servers, and tool-selection accuracy degrades measurably when a lot of tool schema is loaded at once. Deferred loading fetches only the tools a skill actually names.

Skills

Skill	Cadence	Job
fv-onboard	on install	Zero to first attributed data — signup, tracker installed in your repo, Stripe, ads
fv-login	on demand	Connect this machine — browser login, read-only key, no pasting
fv-capabilities	on demand	What's usable right now, and what to connect next
fv-data-health	precondition	Is the data trustworthy? Runs inside every other skill
fv-verify-revenue-feedback-loop	weekly	Is closed revenue actually reaching Google/Meta/LinkedIn?
fv-cut-wasted-spend	weekly	Terms that spend money and produce zero payers → negative keywords
fv-find-leaky-pages	weekly	Pages ranked by revenue lost, not bounce rate
fv-fix-page	on demand	Applies the fix — GitHub PR or Webflow write
fv-design-page-test	on demand	Can this page power an A/B test? Usually no, with the numbers
fv-find-keyword-gaps	monthly	Queries that impress but never pay, ranked by revenue
fv-fix-striking-distance	monthly	Positions 5–20 where a snippet rewrite moves real money
fv-win-back-churned	monthly	Lapsed customers worth recovering → a staged Brevo list
fv-build-audience	on demand	Sized, floor-checked, consent-gated segments (incl. retargeting)

Why there are no "specialist" agents

A reasonable first instinct is to ship a Google Ads specialist, a Meta specialist, a LinkedIn specialist, an SEO specialist and a strategist. This plugin deliberately does not.

A subagent buys you exactly two things: an isolated context window and a restricted tool set. It does not buy expertise — that comes from the instructions, which a skill delivers into the main conversation where the user actually is. Subagents start with no conversation history and cannot ask the user a question mid-run.

Splitting by ad platform is worse than merely unnecessary. The question that matters is cross-platform — which channel gets the next €10k? — and answering it requires one context holding all three platforms on comparable attribution windows. That comparison is the entire reason FullVision exists. Three per-platform agents would each answer confidently from a third of the evidence.

So: platform knowledge lives in shared/platforms/*.md and is read by whichever skill needs it. Jobs are skills. The only subagent is fv-analyst, and it earns its place on context isolation alone — deep multi-view digs whose raw rows nobody needs to see.

How it behaves

It never writes in the turn it analyses. Every write skill proposes a change-list with the evidence inline — spend, payers, revenue, n — then stops and waits for you. You can overrule one line without re-running anything.