Turn Claude Into a Sellable Service: The MCP Workforce Playbook

Claude will plan a content campaign in detail, then stop. It has no hands to run it — it can't reach your video editor, your lead database, or the page you need scraped. MCPs are the hands. The Model Context Protocol (MCP) is an open standard (created by Anthropic) that lets an agent call external tools; a 'connector' is one MCP server wired into Claude. Add the right ones and 'write me an ad' becomes 'generate the ads, edit the video, find the leads, draft the outreach' — end to end. The add flow is the same every time, so learn it once.

1. In Claude, click the + under the message box → Connectors → Manage connectors.
2. Click Add custom connector (if you don't see it, enable Developer Mode in Settings first — the option is hidden without it).
3. Paste the server's remote MCP URL (always https://…, with the right path). Name it.
4. Some servers want auth: an OAuth login, or an API key pasted in Advanced settings — depends on the server.
5. Set the tools to Always allow so the agent isn't stopped for a permission click on every call, then refresh the tools list. The new tools now show up in the connector menu.

Each MCP below is a real, currently-available connector. Pick by the job. URLs and auth are the verified specifics — don't trust spellings you hear in a video.

The moment from the video: the agent generated a folder of video ads, then sorted them itself — clean ones into an approved folder, the ones with garbled on-screen text into rejected. Nobody wrote that rule by hand. It works because a capable model (here, Claude Fable 5) can look at its own output and judge it, then loop until output passes its own review. You reproduce it with one instruction baked into the prompt.

1. Tell the agent the bar in plain language: 'After generating each creative, look at it. Any with misspelled or garbled on-screen text, broken layout, or wrong logo → move to rejected/. Clean, on-brand ones → approved/.'
2. Have it organise into two folders so you review a clean shortlist, not the raw dump.
3. Ask it to keep generating until N approved creatives exist — that's the 'loop until it passes its own review' part.
4. You still review approved/ before anything ships. Model self-QA is best-effort taste, not guaranteed accuracy — it catches obvious junk, not every brand nuance.

One connector is a party trick. Chain them around a real deliverable and you have something to invoice. Knowing the roster isn't the service — the finished, sorted folder you hand back from a brief is. Compose the roster into one prompt with a clear goal and a handoff. Start with ONE of the three packs below, not all three at once. Wiring four MCPs in a single session means juggling four auth flows, rate limits, and model queues — that's a demo, not a repeatable service. Get one pack running end-to-end on a real brief first; add the next once the first is boring.

• Creative pack: Firecrawl scrapes the brand's page → Higgsfield generates image + video ads from it → agent self-sorts into approved/ / rejected/ → you hand over the approved folder.
• Edited explainer: rough footage in → HyperFrames adds captions, chapter titles, lower-thirds and overlays → out comes a postable cut.
• Lead + outreach pack: Clay finds N target creators/companies → Firecrawl deep-researches each → agent drafts a short, specific message per lead referencing their real work → you get a ready-to-review outreach list. Heads-up: Clay's MCP is read-only (search + pull, not bulk enrichment) — if you need columns of enriched data at scale, that step still runs in Clay's own UI. Scope this offer as 'researched leads with drafted outreach,' not 'a full enrichment pipeline.'
• The deliverable is a folder a client can open, not a chat transcript. That's what makes it feel like a service, not a demo.

The lead-research pack is the highest-value and the highest-risk piece. The agent can find people and draft messages beautifully. It must not send them. Drafted ≠ sent.

1. Have the agent produce drafts into a file or sheet — never wire it to actually email or DM.
2. Add a checker step: a second pass that verifies each draft references the person's real work, not a hallucinated detail.
3. You read every message before it goes out. One confidently-wrong line to a real prospect costs more than the whole list is worth.
4. Respect each platform's rules on outreach — automated cold DM is banned in places. Personalised, human-reviewed messages you send yourself are a different thing.

Every sellable version of this follows the same four beats. Write them down once per offer and you have a service you can quote.

1. Scope — name the one outcome you sell ('10 approved ad creatives from your brand page', 'a captioned 60-sec explainer', '25 researched leads with drafted outreach'). Specific and checkable beats 'AI marketing help'.
2. Goal — give the agent the outcome up front in one clear instruction, plus the self-QA bar. A well-specified goal is what lets it run without you babysitting.
3. Workflow — list which MCPs in which order (scrape → generate → sort, or find → research → draft). This is your reusable recipe.
4. Handoff — define the artifact: a named folder, a sheet, a file. State explicitly what you deliver and what stays human (the send button, the final approve).