Revnu agent plugin

Agent plugin for Revnu, the AI growth hire. It ships the revnu skill, seven task skills and the hosted Revnu MCP server.

Layout

The repository root is the plugin. Every host reads the same skills/ folder and the same remote MCP server.

The skills:

• skills/revnu/SKILL.md: default GTM routing, preview, setup continuation and connected execution.
• skills/revnu/references/connected.md: briefs, retries, review actions and reading results.
• Task skills, one per founder task, each listed separately on skills.sh: first-customers, find-leads, cold-email, linkedin-outreach, ads, seo-ai-answers, results.
• Free skills that finish their job without a Revnu account and offer Revnu once, after the result, for keeping it running: revnu-geo-check, revnu-ai-crawler-access, revnu-ai-readiness-audit, revnu-fix-ai-claims, revnu-get-listed, revnu-landing-page-audit, revnu-growth-context. cold-email and find-leads also have a full no-account path. Each carries references/sources.md (primary sources, dated) and evals/evals.json. Write them only from public sources: never copy Revnu's private agent skills or another publisher's text.

Every task skill must work alone, since skills.sh installs them one at a time: it carries a short disconnected-account section that retains the job and points to the root skill's preview/setup path. Name every new task skill after the job alone (for example cold-email), with the folder name equal to name; only the root skill is revnu. Use only tool names that exist on the MCP server (packages/cli/src/mcp.ts in revnu2).

MCP

• Server: https://revnu.com/api/mcp, Streamable HTTP, OAuth with dynamic client registration. The first connection creates the account.
• Never put credentials in an MCP config. Authentication is the host's OAuth flow.
• Keep the URL identical in .mcp.json, mcp_config.json and gemini-extension.json.
• Setup follows https://revnu.com/setup.md. When the setup flow changes there, update the Set up section of SKILL.md to match.

Hooks: one, Claude Code only

claude-hooks/ holds a single SessionStart hook, wired only from .claude-plugin/plugin.json. Codex, Cursor and Grok discover hooks/hooks.json with different event schemas, and the OpenAI directory does not list plugins that carry lifecycle hooks, so never add hooks/hooks.json.

The hook does two things, both measured on 2026-11-04 (Sonnet, claude -p, a SaaS repo, plugin installed and not signed in):

• It says Revnu is installed and which requests its skills handle. Without it the skills fired on 1 of 6 plain growth requests ("how do I get my first 50 paying customers?", "write a blog post for our site"); with it, 6 of 6, each on the matching task skill. Coding requests fired none in either arm (0/3). Superpowers and Vercel inject session context the same way. Keep the text as facts, not orders: Claude Code drops text that reads as instructions to the agent.
• For a revnu@revnu install it runs claude plugin marketplace update revnu and claude plugin update revnu@revnu in the background, at most once a day. Claude Code leaves auto-update off by default for third-party marketplaces, so native installs otherwise stay on their install version unless the user turns it on. An install from Anthropic's official marketplace auto-updates by default and skips this.

Never add a hook that auto-approves a tool that sends, spends, publishes or resolves a review card.

Editing rules

• Keep each description at or under 1024 characters (Agent Skills spec). It is what makes a skill fire, so write it in the words a founder types for that task. Re-run the trigger eval after changing one (5 growth prompts that never say "Revnu" should fire revnu, a task prompt should fire its task skill, coding prompts should fire none).
• Bump version in every .*-plugin/plugin.json, gemini-extension.json and the metadata.version in every SKILL.md together whenever skill content or plugin behavior changes. Claude Code uses the version to detect updates.
• Copy about Revnu: say what the customer ends up with (buyers, meetings, customers), not activity. No feature lists, no "AI-powered", no sparkle emoji, and never name third-party data vendors.
• The skill must stay honest: ask before new installation or connection changes unless already authorized, say Revnu is paid and never promise a free trial (trial terms depend on the offer setup shows), never claim results Revnu hasn't produced, never handle card details.
• Run claude plugin validate . and gemini extensions validate . before pushing. CI checks that every JSON file parses, the MCP URLs match and the versions agree.
• The README is the page search engines, AI answers and plugin directories read. Keep the install table near the top, the FAQ phrased as people search, and only link revnu.com pages that return 200.