Media forge

Media Forge

You turn a prompt into a file. Image, video, or audio — through Replicate or fal.ai, the two aggregators that front nearly every open and commercial generative model. You are the media half of produce in trend -> remix -> produce -> post -> analytic.

Pure worker: no state, no SDK, no persisted secret. You read one token from the environment, submit a job, wait for it, download the outputs, and print MediaArtifact JSON. The orchestrator resolves the token (gf token replicate|fal), runs you, and attaches your artifacts to a ContentItem through the seam.

---

Run It

bun run scripts/forge.ts \
  --provider replicate|fal \
  --modality image|video|audio \
  --model <model-id> \
  --prompt "..." \
  [--input '{"aspect_ratio":"16:9","num_outputs":2}'] \
  [--out .genfeed/artifacts]

Requires Bun 1.1+. Zero dependencies — only Node built-ins and global fetch.

--input is merged over {prompt} and passed verbatim as the model's input object, so any model-specific parameter is reachable without code changes.

---

Providers

Replicate

• Token: REPLICATE_API_TOKEN.
• Uses the unified POST /v1/predictions endpoint, which accepts owner/name, owner/name:version, or a bare version id as --model.
• Sends Prefer: wait for synchronous completion, then polls urls.get for anything still running.

bun run scripts/forge.ts --provider replicate --modality image \
  --model black-forest-labs/flux-1.1-pro --prompt "a t-rex on a skateboard, cinematic"

fal.ai

• Token: FAL_KEY.
• Uses the async queue: POST https://queue.fal.run/<model>, polls the returned status_url until COMPLETED, then fetches response_url.

bun run scripts/forge.ts --provider fal --modality video \
  --model fal-ai/ltx-video --prompt "neon city flyover at night"

Both providers cap waiting at ~10 minutes (240 polls × 2.5s), enough for slow video jobs.

---

Output

Outputs are downloaded into --out (default .genfeed/artifacts), one file per result, named art_<uuid>.<ext> (extension inferred from the URL, else png/mp4/mp3 by modality). stdout is:

{
  "count": 1,
  "artifacts": [
    {
      "id": "art_2f1c...",
      "modality": "image",
      "provider": "replicate",
      "model": "black-forest-labs/flux-1.1-pro",
      "prompt": "a t-rex on a skateboard, cinematic",
      "path": ".genfeed/artifacts/art_2f1c....png",
      "meta": { "sourceUrl": "https://...", "index": 0, "createdAt": "..." }
    }
  ]
}

Each artifact matches the MediaArtifact shape in genfeed-connector/lib/schema.ts, so the orchestrator can push it straight onto ContentItem.artifacts[].

Output URLs are extracted by walking the provider's response JSON and collecting every http(s) URL, so it works across model output shapes (string, array, {images:[{url}]}, {video:{url}}, {audio:{url}}, …) without per-model branching.

---

How It Fits The Loop

# orchestrator resolves a scoped token through the seam, never seeing the raw vault credential
export REPLICATE_API_TOKEN="$(bun run ../genfeed-connector/gf.ts token replicate)"

# forge a hero image for an item, capture artifacts
ART=$(bun run scripts/forge.ts --provider replicate --modality image \
  --model black-forest-labs/flux-1.1-pro --prompt "$THESIS")

# orchestrator merges $ART.artifacts into the ContentItem and advances the stage

---

Security

• Tokens (REPLICATE_API_TOKEN, FAL_KEY) are read from the environment, used in memory, and never written to any file. Resolve them through the connector so the long-lived credential stays in the env or the genfeed vault.
• The only outbound calls are to the chosen provider's API and the output-file URLs it returns.
• Generated files are written only under --out; nothing else touches the filesystem.