ElevenLabs' hosted MCP puts over fifty image, video, voice and music models inside Claude through one OAuth sign-in, no API keys. These are the five client pipelines I would wire through it first, each with the exact brief you paste into Claude, plus the cost gate that keeps a run from spending before you approve it.
In Claude Code, run the one-line install below (it opens an ElevenLabs sign-in in your browser, that is the whole setup). In Claude desktop or web, open Settings, Connectors, search ElevenLabs and press Connect. Then ask Claude: list the node types and models my ElevenLabs workspace can run. It answers with the real roster for your account (mine returned 25 image models, 47 video models, 5 voice models, music, sound effects and Claude itself as a text node). Everything below is a plain-English brief you paste into that same chat.
02
02Pipeline 01
Product ad: one still, then Omni animates it
This is the flow that made the clip in the reel: a Nano Banana Pro still, then Gemini Omni Flash turns it into a five-second product shot. Paste the brief in the code box, swap the product line for the client's. Two rules that make it look shot, not generated: describe the imperfections (dust in the light, a fingerprint on the glass, fabric grain) and give the video step motion only, never a second description of the scene. Ask for two stills, pick one, animate one. Deliverable: a 16:9 or 9:16 clip the client can run as a paid ad.
03
03Pipeline 02
Voice-over in the client's own voice, mixed and cut
Ask Claude to list my voices and pick the client's cloned voice (or a stock one for the draft). Brief: make a 30-second voice-over from this script on voice X, a 30-second instrumental bed that stays under speech, and one sound effect for the logo hit, then compose them on a timeline with the voice on top and the music at 20 percent. Claude builds a TTS node, a music node, a sound-effects node and a composition node, wires them and runs the whole thing. You approve the cost once. Deliverable: one mixed audio file for the client's reel or podcast intro.
04
04Pipeline 03
Talking head from one photo
For clients who will not go on camera. Brief: use this headshot and this script: generate the speech on voice X, then animate the photo into a talking-head video with OmniHuman 1.5 (or Creatify Aurora for a studio look), 9:16. Upload the headshot when Claude asks, or point it at an asset already in the workspace library. Check the hands and the mouth at 1:1 before you send anything. Deliverable: a UGC-style explainer without a shoot day.
05
05Pipeline 04
The voice agent, described in a sentence
The same connector manages ElevenAgents. Brief: create a voice agent for a dental clinic that books and reschedules appointments, warm and brief, first message 'Hi, this is the clinic assistant, how can I help', voice X, then give me the share link. Claude creates it, sets the system prompt, voice and first message, and returns the widget link you send the client to test. Ask it to estimate this agent's monthly LLM cost at 300 calls before you quote. Deliverable: a testable agent the same afternoon.
06
06Pipeline 05
The cost gate: nothing spends until you say so
Every generation tool has an estimate mode. Make it a habit: estimate first, do not run at the end of any brief, read the credit number, then say run. Keep write access on the agents side restricted to the people who need it (deleting an agent is destructive and the docs say so), and review every tool call in the client before approving. On my side the rule is one flow per client job, named after the job, so the canvas link goes straight into the delivery folder.
Install line + the Pipeline 01 brief · paste into Claude
# Claude Code · one line, then sign in with ElevenLabs in the browser claude mcp add --transport http elevenlabs https://api.elevenlabs.io/v1/mcp
# Pipeline 01 · paste this in the chat Create a flow called CLIENT product ad. Step 1: a Nano Banana Pro still, 16:9, 2K: ultra-realistic candid product photograph of [PRODUCT] on [SURFACE], soft window daylight from the left, 85mm macro, shallow depth of field, dust in the light, a faint fingerprint on the glass, fabric grain, muted natural color, subtle film grain, no text, no logo, no CGI look. Make 2, I will pick one. Step 2: animate the pick with Gemini Omni Flash, 5 seconds, 720p: slow dolly-in, dust drifting in the light, a soft shadow crossing as a cloud passes. Everything else stays exactly the same. Estimate first, do not run.
Want this running for you?
Build it step by step inside AI Creative OS, with me in the thread.
