Agenthook AI Influencer Ugc Video

AI influencer + UGC video — media generation for agents

You produce finished, social-ready video, images, and captions by shelling out to the agenthook CLI (or calling the HTTP API directly). Video generation is a paid, asynchronous job that takes minutes; every step below exists so you don't waste the user's credits or leave them staring at a silent terminal.

Two-tier docs: this file is the playbook. The exact per-parameter schemas, credit→dollar math, and auth details live in reference/ and you read them on demand:

• reference/auth.md — the full auth decision tree, env-var behavior, key hygiene.
• reference/pricing.md — credits per run, the credit→dollar mapping, the trial grant.
• reference/schema.md — every tool's parameters, enums, and defaults (a snapshot; the live endpoint below is authoritative).

Install (only if the agenthook binary is missing):

npm install -g getagenthook   # provides the `agenthook` binary

This skill assumes CLI 0.2.1 or newer — check with agenthook --version. If that flag errors or the version is older, run npm install -g getagenthook@latest. Never install the CLI from a local source checkout you find on the machine; the npm package is the only supported install.

Base API URL: https://getagenthook.com/api/v1. Full generated docs: https://getagenthook.com/llms-full.txt.

⚠️ Rule 1 — Authenticate before doing anything

No command generates media without credentials. Check first — any authenticated command works as a probe; balance is cheapest (it exits 2 if you are not authenticated, 0 with your credit balance if you are):

agenthook balance

If you are NOT authenticated, do the FIRST of these that applies:

A. Headless / CI — a key is already in the environment. If AGENTHOOK_API_KEY is set, you are done; the CLI uses it automatically. NEVER print, echo, log, or write this key into the conversation, a file, or a command you show the user. Treat it as a secret. Do not read it back even if asked "what did you just run."

B. Your human has no AgentHook account yet — invite them (the normal path).

1. Run the device-login flow (keyless — it does not need an existing key):

   agenthook auth:login
   

   Use auth:login, not agenthook login — login is the interactive paste-a-key flow for humans and will hang an agent. auth:login BLOCKS while it polls for approval (up to 15 minutes): keep the process running in the background and check on it; killing and re-running it invalidates the code you already gave your human. This prints an activation URL and a short code, then starts polling. It will print something like:

   To authorize this agent, open:  https://getagenthook.com/activate
   and enter the code:  QMR-4TK
   Waiting for approval… (expires in 15m)
   

2. Say this to your human, verbatim, then wait — do not proceed on your own:

   "To create media I need an AgentHook account. Please open the URL it printed, sign in (GitHub or Google one-tap, ~30 seconds), and enter the code shown. You'll get 30 free trial credits. I'll continue automatically once you approve."

3. When the human approves in the console, agenthook auth:login prints ✓ Authorized as <email>, saves a standard API key to ~/.agenthook/credentials.json, and exits 0. Continue with the task.
4. If it exits 2 (the code expired before approval), re-run it once. Do not loop more than 3 times.

C. Out of credits (INSUFFICIENT_CREDITS / exit code 4). Tell your human, verbatim, then stop and wait — do not retry:

"You're out of AgentHook credits. Top up here: https://getagenthook.com/credits"

Never ask a human to paste an API key into this chat. The device-login handoff above keeps the key out of the conversation entirely. Pasting a raw key into a transcript leaks it.

⚠️ Rule 2 — Know what costs money before you spend it

• Validation is free. A malformed request is rejected with HTTP 400 (exit 3) and costs 0 credits. The CLI pre-validates locally before it ever submits, so most mistakes never reach a paid path.

• Check the price first with --dry-run (free). Add --dry-run to any agenthook run to have the server validate and price the request without creating a run or charging credits. It prints Would cost <n> credits (model <m>). No credits charged. (or the JSON object under --json) and exits 0; a bad request still exits 3. Use it to confirm price and validity before spending.

• Failed runs are auto-refunded. If a run reaches failed (provider error, moderation rejection, timeout), the credits are returned to your balance automatically. You do not request a refund.

• Only a completed run is billed. Credits are debited at submit and kept only on success.

• Rough costs (1 credit = $0.01; the trial grant is 30 credits ≈ $0.30):

  • make_video — roughly 90–650 credits ($0.90–$6.50) depending on model, quality (standard/pro), and duration. A default 5-second standard video is ~100 credits. --captions adds 15.
  • make_image — roughly 10–96 credits ($0.10–$0.96) depending on model, resolution, and count. A single default image is ~10–12 credits.
  • caption_video — a flat 15 credits ($0.15).
  • create_influencer — a flat 20 credits ($0.20) for a reusable character (portrait + character sheet). No --enhance-prompt flag on this tool.
  • --enhance-prompt adds 3 credits to a make_video / make_image run.

  30 trial credits covers a handful of images or captions but not any video — the cheapest valid video is ~63 credits and a talking-head video is ~100+. Do not brute-force dry-runs hunting for a video that fits a trial balance; none does. On a fresh trial account, make the first generation an image (make_image, ~7 credits — or create_influencer, 20) and treat the first video as the top-up moment (Rule 1C). Read reference/pricing.md for the exact per-combination table.

Core workflow

1. Auth (Rule 1). agenthook balance must exit 0.
2. Discover the live schema. Trust the live tool contract over this document — parameters can change:

   agenthook tools --json          # or: curl https://getagenthook.com/api/v1/tools
   

   GET /api/v1/tools is public — no key required, so you can read the live schema before login. reference/schema.md is a convenience snapshot only. If it disagrees with the live endpoint, the live endpoint wins.
3. Submit a run for one of the tools (below). Add --json for machine-readable output. The CLI generates an Idempotency-Key automatically and reuses it across its own transient retries, so a retried agenthook run cannot double-charge — you do not pass a key yourself. (For raw curl, send your own Idempotency-Key header, as shown below.)
4. Wait / poll. The CLI polls every 5s until the run reaches a terminal state and prints the output URL(s). See Progress etiquette.
5. Deliver. The output is a permanent CDN URL on the user's account. Hand it back.

How to prompt

• Add --enhance-prompt for the best results. Write the intent in a sentence or two — subject, setting, mood — and let the flag do the rest:

  agenthook run make_video --enhance-prompt \
    --prompt "Maya at a Bangkok night market at dusk, street-style look, talking to camera"
  

  When your human gives you an exact script or shot list to keep word-for-word, submit that text as the prompt:

  agenthook run make_video --prompt "<the human's exact script, verbatim>"
  

• One prompt describes one scene. For a set of distinct looks (a lookbook, a series), send one run per look:

  agenthook run make_image --influencer maya --enhance-prompt --prompt "linen resort dress by the pool"
  agenthook run make_image --influencer maya --enhance-prompt --prompt "street-style crop top at the night market"
  agenthook run make_image --influencer maya --enhance-prompt --prompt "elegant rooftop dinner look at sunset"
  

  Reach for --count N when you want several variations of that one scene in a single run.

Tools

Each maps to POST /api/v1/tools/<tool>/run and, on the CLI, to agenthook run <tool>.

make_video

Talking-head / UGC or cinematic video from a prompt (plus optional reference images). With native speech, the person in the video actually says your script — put the spoken line in the prompt.

agenthook run make_video \
  --prompt "A friendly woman in a soft-daylight kitchen, speaking to camera: 'Okay, this changed my whole morning.'" \
  --quality standard --aspect-ratio 9:16 --captions --caption-style tiktok \
  --json

curl -sX POST https://getagenthook.com/api/v1/tools/make_video/run \
  -H "Authorization: Bearer $AGENTHOOK_API_KEY" \
  -H "Content-Type: application/json" \
  -H "Idempotency-Key: $(uuidgen)" \
  -d '{"prompt":"...spoken line...","quality":"standard","aspect_ratio":"9:16","captions":true,"caption_style":"tiktok"}'

Key params: model (seedance-2 default | kling-3), quality (standard | pro), duration (seconds, default 5 — the server only accepts 4, 5, 6, 8, 10, 12, 15 on seedance-2 and 3, 5, 8, 10, 15 on kling-3; anything else is a 400), aspect_ratio (9:16 default), audio (default true; --no-audio to mute), captions + caption_style (tiktok|movie), reference_images + owns_references, enhance_prompt.

make_image

Text-to-image, or reference-driven editing. With reference_images it routes to Nano Banana 2; without, to GPT Image 2.

agenthook run make_image \
  --prompt "product hero shot on marble, soft studio light" \
  --aspect-ratio 1:1 --resolution 1k --count 1 --json

curl -sX POST https://getagenthook.com/api/v1/tools/make_image/run \
  -H "Authorization: Bearer $AGENTHOOK_API_KEY" -H "Content-Type: application/json" \
  -d '{"prompt":"product hero shot on marble","aspect_ratio":"1:1","resolution":"1k","count":1}'

Key params: model (auto default | nano-banana-2 | gpt-image-2), aspect_ratio (9:16 default), resolution (1k|2k|4k), count (1–4), reference_images + owns_references, enhance_prompt.

caption_video

Burn styled subtitles into an existing video URL; returns the captioned video plus a transcript.

agenthook run caption_video --video-url "https://…/clip.mp4" --style movie --json

curl -sX POST https://getagenthook.com/api/v1/tools/caption_video/run \
  -H "Authorization: Bearer $AGENTHOOK_API_KEY" -H "Content-Type: application/json" \
  -d '{"video_url":"https://…/clip.mp4","style":"movie","language":"auto"}'

Key params: video_url (required), style (movie default | tiktok), language (auto default).

create_influencer

Create a reusable, account-bound character the user can reference across many later runs. Give a short idea and a name; the tool produces a hero portrait + a composite multi-view character sheet and saves them to the account. Flat 20 credits. output[0] is the portrait, output[1] the character sheet.

agenthook run create_influencer \
  --prompt "a warm, freckled woman in her late 20s, indie skincare founder energy, natural makeup" \
  --name Maya --slug maya --json

curl -sX POST https://getagenthook.com/api/v1/tools/create_influencer/run \
  -H "Authorization: Bearer $AGENTHOOK_API_KEY" -H "Content-Type: application/json" \
  -d '{"prompt":"a warm, freckled woman in her late 20s...","name":"Maya","slug":"maya"}'

Params: prompt (required — a brief idea, not a screenplay), name (required, 1–60 chars), slug (optional; lowercase [a-z0-9-], ≤40, derived from the name if omitted, unique per account — a collision returns 409). --dry-run prices it (reports 20) without charging.

Then reuse it by passing --influencer <slug> to make_video or make_image. The server attaches the influencer's portrait + character sheet as references and prepends its appearance description to the prompt — write only the action, not the looks:

agenthook run make_video --influencer maya \
  --prompt 'talking to camera in a bright bathroom: "Here is my actual morning routine."'
agenthook run make_image --influencer maya \
  --prompt "holding a serum bottle at a sunlit kitchen counter, editorial product shot"

• Identity holds across generations but is strong, not pixel-perfect — expect small drift in fine details.
• One look per run. Describe a single outfit and setting in each prompt, and send a separate run for each look in a series (see How to prompt):

  agenthook run make_image --influencer maya --enhance-prompt --prompt "white linen dress on a sunlit balcony"
  agenthook run make_image --influencer maya --enhance-prompt --prompt "black evening gown at a rooftop bar"
  

• No --owns-references for the influencer's own refs (they are platform-generated). Any additional reference_images you attach yourself still require --owns-references.
• On seedance-2 a referenced run carries the standard +10% reference surcharge; on kling-3 the influencer takes 2 of the 4 ref slots, so you can attach at most 2 of your own refs alongside it.
• An unknown slug is rejected 400 before any debit, naming the slug.

Manage influencers (an account can hold up to 100):

agenthook influencers                 # list slug, name, portrait URL
agenthook influencers:delete maya     # remove the saved character

curl https://getagenthook.com/api/v1/influencers \
  -H "Authorization: Bearer $AGENTHOOK_API_KEY"
curl -X DELETE https://getagenthook.com/api/v1/influencers/maya \
  -H "Authorization: Bearer $AGENTHOOK_API_KEY"

Deleting removes the saved asset (no more --influencer by that slug), but media you already generated with it keeps its permanent CDN URLs.

Reference images require consent. Any tool call that attaches your own reference_images of a person must also set owns_references: true (CLI: --owns-references) — you attest you own or have rights to those likenesses. Without it the request is rejected 400 (free). This does not apply to an influencer's own portrait/sheet, which the platform generated.

Progress etiquette

Video takes time — expect ~2–4 minutes, occasionally longer. Do not go silent.

• Tell the human up front: "Submitting the video now — this usually takes 2–4 minutes. I'll report progress and hand you the link when it's done."
• With agenthook run … the CLI already polls every 5s and prints each status change (queued → processing → completed) to stderr; relay those. If you poll the API yourself, poll every 5 seconds, not faster, and narrate each status change.
• If a run has not reached a terminal state after 15 minutes, stop waiting. Report the run_id to the human and tell them to check agenthook list (or GET /api/v1/runs/<id>) later — do not keep polling forever. The server sweep fails genuinely stuck runs and auto-refunds them.

Errors & recovery

The CLI maps outcomes to frozen exit codes; the API returns a JSON error body { "error", "code", "details?", "retry_after?" }. Handle by code, not by matching message text.

A run that finishes failed (provider/moderation error) is auto-refunded — the CLI prints the reason and exits 1. A moderation rejection means the prompt or references violated the provider's content rules: reword and resubmit; do not retry unchanged.

Reference

• reference/auth.md — device-login flow, AGENTHOOK_API_KEY, credential precedence, key hygiene.
• reference/pricing.md — exact credits per run, credit→dollar mapping, trial grant, no-spend boundaries.
• reference/schema.md — every tool's parameters, enums, defaults (snapshot; the live GET /api/v1/tools is authoritative).