Hermoso — the marketing MCP

Hermoso is an AI ad studio an agent operates end to end: research the ads already winning in any market, create finished on-brand image and video ads (including UGC-style talking-head spots with native speech), publish or schedule to the user's own channels, and build and read ad campaigns on eleven ad platforms. 366 tools over one MCP server; every area works on its own — nothing requires another step first.

Connect (once)

1. The user creates a free account at https://app.hermoso.ai and makes an agent key under Settings → Agents & API (it starts with hmk_). Free accounts start with credits; renders are quoted before they run.
2. Add Hermoso as an MCP server in OpenClaw (Settings → MCP → Add server), either transport:
   • Streamable HTTP: URL https://app.hermoso.ai/mcp, header Authorization: Bearer hmk_…
   • Stdio: command npx, args -y hermoso mcp, environment HERMOSO_TOKEN=hmk_…
3. Verify it answers: openclaw mcp doctor hermoso --probe. Then call hermoso_capabilities (free) once — it returns the live model catalog with exact credit costs, which is what to quote before any render.

The tools, by job

• Research: find_competitors, pull_competitor_ads, research_ads, search_meta_ads, search_google_ads, search_linkedin_ads, search_tiktok, search_instagram, search_youtube, mine_angles, analyze_video. Nearly free. Report the strongest hooks and angles and quote the real headlines.
• Create: draft_brand (onboard a brand from its website once) → plan_ad → render_ad (the Studio pipeline: brand, product photo, storyboard, native audio, read-back of the delivered file). generate_image / generate_video for raw model runs, make_template_ad for native formats, make_explainer for narrated explainers. Video renders take 2–6 minutes: render_ad returns a job id — keep calling get_job until it reports done, and never describe a video before that.
• Publish & schedule: post_to_meta (Facebook + Instagram + Threads), post_to_x, post_to_linkedin, post_to_tiktok, post_to_youtube, post_to_pinterest, post_to_reddit, post_to_bluesky, post_to_telegram, post_to_google_business; schedule_post, list_scheduled, reschedule_post, cancel_scheduled. upload_file turns any local or external file into a URL these accept. Channels are connected by the user in the app (OAuth needs a browser); list_connectors says which are live.
• Ads management: create_meta_campaign / _adset / _ad, create_google_ads_campaign / _ad_group / _ad, and the TikTok, LinkedIn, Pinterest, Reddit, Microsoft, X and ChatGPT equivalents; meta_insights, google_ads_report and the per-platform reports. Everything is created paused and read back from the platform before it is described. Only the set_*_ads_status tools with confirm:true can make money move — never pass confirm unless the user asked to go live.

Shell shortcuts (same account, via exec)

npx -y hermoso auth login --token hmk_…, then: hermoso competitors <domain>, hermoso ads pull --company "<name>", hermoso research "<request>", hermoso brand draft --domain <domain>, hermoso create --brand <name> --product "<product>" --format video, hermoso generate image --prompt "…", hermoso jobs get <id> --wait, hermoso credits, hermoso capabilities. Add --json for machine output.

Rules the agent should keep

• Quote credits (from hermoso_capabilities) before a render or a campaign build; the user pays per render, not per seat.
• A render is finished when its URL comes back — not before.
• If a tool answers "not connected", ask the user to connect that channel in the app; the agent cannot complete OAuth.
• Nothing here spends ad budget without the explicit confirm flag.