Using from Claude Code & Cursor

Hand the CLI to your coding agent — it’s built for this.

AgentHook is designed to be driven by agents, not humans in a dashboard. The CLI is the agent surface today: any agent that can run shell commands (Claude Code, Cursor, Codex-style agents, n8n exec nodes) can generate media with it. An MCP server and an installable skill pack are coming later — both will be thin wrappers over the same API.

Setup

Install and authenticate once on the machine your agent runs on:

npm i -g getagenthook
agenthook login   # paste an API key from the console

(Keys are created in the console — see Authentication.)

Tell your agent about it

Add a note to your project’s agent instructions (CLAUDE.md, Cursor rules, etc.):

## Media generation

The `agenthook` CLI generates videos and images (real credits — don't run speculatively):

- `agenthook tools` — print every tool and its full input schema
- `agenthook run make_video --prompt "..."` — video; put spoken lines in the
  prompt ('She says: "..."'); flags: --captions, --enhance-prompt; with reference
  images, --owns-references is required (likeness consent)
- `agenthook run make_image --prompt "..."` — images (refs = editing)
- `agenthook run caption_video --video-url <url>` — burn in subtitles
- `agenthook list --search "<keywords>"` — find past generations before
  making new ones
- `agenthook balance` — check credits first

Runs print a durable media URL when done.

Why this works well for agents

• Self-describing schemas. agenthook tools (backed by GET /api/v1/tools) prints every argument, type, default, and enum — the agent never has to guess or memorize the API.
• Local pre-validation. The CLI checks everything deterministically checkable — enum values, prompt length caps, references without the owns_references attestation — before any network call, and fails with a message the agent can read and self-correct from. No credits are spent on malformed requests.
• Submit–poll handled for you. run blocks until the media is ready and prints the URL — one command, one result, no polling loop for the agent to write.
• Searchable memory. agenthook list --search "..." full-text-searches past runs (prompt, enhanced prompt, transcript), so an agent can reuse yesterday’s video instead of paying to regenerate it.

No CLI? Use HTTP directly

Everything the CLI does is plain REST — agents that only speak HTTP can hit the API directly:

curl -X POST https://getagenthook.com/api/v1/tools/make_image/run \
  -H "Authorization: Bearer $AGENTHOOK_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"prompt": "Flat-lay of a lime-green running shoe on dark slate", "aspect_ratio": "1:1"}'

Then poll GET /api/v1/runs/{run_id} until completed. See the Quickstart for the full flow.

Coming later

• MCP server — the same three tools over the Model Context Protocol, for MCP-native clients.
• Skill pack — installable agent skills that wrap the CLI with usage guidance.

Both will read their schemas from GET /api/v1/tools, same as the CLI — one schema source, no drift.