run-noodle-action

Run a saved nanoodle workflow in your CI pipeline. Design a graph visually in the nanoodle editor, commit the noodle-graph.json, and this GitHub Action runs it against the NanoGPT API — generate the release jingle, the OG image, or the changelog art right from a workflow.

Wraps the zero-dependency nanoodle CLI (npx nanoodle run), pinned to an exact version.

Over a bare npx nanoodle run step, this action passes your API key to the CLI only via the child process environment (never argv or logs), pins the CLI to an exact version you control, and exposes typed step outputs (cost-usd, out-dir, result-json). Not in CI? Run graphs directly with nanoodle-js (Node) or nanoodle-py (Python).

This spends real money. Every run bills your NanoGPT balance per generation. Trigger it from workflow_dispatch or a release tag — never on every push.

Usage

name: release art
on:
  workflow_dispatch:      # manual button — recommended (each run costs money)

jobs:
  generate:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v4

      - id: noodle
        uses: nanoodlecom/run-noodle-action@v1
        with:
          graph: art/noodle-graph.json
          api-key: ${{ secrets.NANOGPT_API_KEY }}
          inputs: |
            Text=release ${{ github.ref_name }} poster, bold, celebratory

      - uses: actions/upload-artifact@v4
        with:
          name: release-art
          path: ${{ steps.noodle.outputs.out-dir }}

See .github/workflows/demo.yml for a runnable example (workflow_dispatch only, for the same reason).

Run straight from a share link

No file to commit — paste the link the editor's Share button gives you into graph: and CI runs that pipeline. A full URL, a short link, or a bare #g=/#j=/#a= fragment all work (needs nanoodle-version >= 0.2.0).

      - id: noodle
        uses: nanoodlecom/run-noodle-action@v1
        with:
          graph: https://nanoodle.com/#g=H4sIAAAA...   # your share link
          api-key: ${{ secrets.NANOGPT_API_KEY }}
          inputs: |
            Text=release ${{ github.ref_name }} poster, bold, celebratory

Direct #g=/#j=/#a= links decode locally with no extra network hop; da.gd/TinyURL short links are followed to find the underlying fragment.

The flow

1. Design and test your workflow at nanoodle.com — pick your models there too.
2. Hit save; download noodle-graph.json.
3. Commit the JSON to your repo.
4. Point this action at it. npx nanoodle inspect your-graph.json (offline, free) shows the input and setting keys the graph accepts.

Secrets setup

Add your NanoGPT API key as a repository secret named NANOGPT_API_KEY (repo → Settings → Secrets and variables → Actions). Pass it via api-key: ${{ secrets.NANOGPT_API_KEY }}. The action hands the key to the CLI through an environment variable only — it never appears in the command line or the logs.

Inputs

input	required	default	description
graph	yes	—	The graph to run — either a path to a saved noodle-graph.json in your repo, or a nanoodle share link (a full https:// share URL, a da.gd/TinyURL short link, or a bare #g=/#j=/#a= fragment). Requires nanoodle-version >= 0.2.0.
api-key	yes	—	NanoGPT API key — pass ${{ secrets.NANOGPT_API_KEY }}.
inputs	no	""	Workflow inputs, one KEY=VALUE per line. Blank lines and # comments are skipped. @path values read files, same as the CLI.
set	no	""	Setting overrides, one node.setting=value per line (e.g. n3.model=flux-dev).
out-dir	no	nanoodle-out	Where media outputs and the result JSON are written.
timeout-ms	no	—	Overall run timeout in milliseconds.
nanoodle-version	no	0.4.0	Exact CLI version to run — pinned, never floating. Share-link graphs need >= 0.2.0; local media nodes need >= 0.4.0.

Outputs

output	description
out-dir	Directory with saved media outputs plus nanoodle-result.json.
cost-usd	Total run cost in USD as reported by NanoGPT (a floor if any call omitted a price).
result-json	Path to the machine-readable result: outputs, cost, per-node statuses.

Media outputs are saved into out-dir named after their output key (Image.jpg, Song.mp3, ...). Text outputs appear in the result JSON.

Limitations

• Feed-forward DAGs only — that is all nanoodle graphs are; there are no loops or agents.
• Some local media nodes need ffmpeg. With the default nanoodle 0.4.0, local media nodes (resize, vframes, combine, soundtrack, trim, extractaudio) run headlessly: the CLI prefers a pure-JS path that matches the browser (lossless mp4 remux, PCM-WAV trim, PNG resize) and falls back to ffmpeg on PATH for everything else. ffmpeg is a soft dependency — if a node needs it and it's missing, the CLI fails with a clear error. On runners without it, add a step like sudo apt-get install -y ffmpeg before this action. See the nanoodle-js docs.
• Graphs must have models set. The editor writes the models you chose into the JSON when you save — pick models in the editor before downloading, or override with set: (e.g. n3.model=flux-dev).
• Each run spends NanoGPT balance; a failed run may still have spent on the nodes that completed (partial results and per-node costs are in the result JSON, and the step fails).

Marketplace

Publishing to the GitHub Marketplace requires this repository to be public plus a v1 release/tag. The v1 tag is pushed; the Marketplace listing itself is a post-public step — not done yet.

License

MIT — see LICENSE. Not affiliated with NanoGPT. Build workflows at nanoodle.com.