Using GitHub Models with your GitHub Workflows

GitHub Models are AI capabilities built directly into GitHub, acting like an “AI lab” within your repository. They let you experiment, compare, and run models seamlessly as part of your existing workflow – no switching tools required.

Key benefits include:

• Native integration – call AI models directly from workflows
• Quick experimentation – use the playground to refine prompts and compare outputs
• Reusable automation – version, reuse, and scale workflows across teams
• Infrastructure-free – GitHub handles hosting, permissions, and authentication

This makes it much easier to move from experimentation to production AI without context switching or security headaches.

What Is AI Inference?

In simple terms, inference is running a trained AI model against new input to get an output.

• Training is teaching the model
• Inference is asking it questions

Example:

• Input: “Summarise this pull request.”
• Output: “This PR updates the settings page on getting started.”

In your workflows, inference means you can feed code, text, or structured data into a model—and instantly use the results to drive automation.

Calling GitHub Models with actions/ai-inference

The actions/ai-inference GitHub Action gives you a standard way to run inference in a workflow job.

What it supports:

• Inline prompts (directly in workflow YAML)
• .prompt.yml files for reusable, structured prompts
• Variable templating (pass runtime data into prompts)
• JSON schema validation (enforce safe, structured outputs)

Required: set workflow permissions:

permissions:
  models: read

Without this, your workflow won’t have access to GitHub Models.

Example: Summarising a README

Here’s a simple first step: automatically summarise your README when the workflow runs.

name: 'AI inference: Summarise README'
on: workflow_dispatch

jobs:
  inference:
    permissions:
      models: read
      contents: read
    runs-on: ubuntu-latest
    steps:
      - name: Checkout repository
        uses: actions/checkout@v4

      - name: Read README content
        id: readme
        run: |
          content="$(cat README.md)" # Read entire README
          echo "content<> $GITHUB_OUTPUT
          echo "$content" >> $GITHUB_OUTPUT
          echo "EOF" >> $GITHUB_OUTPUT

      - name: Summarise README with AI
        id: inference
        uses: actions/ai-inference@v1
        with:
          model: openai/gpt-4.1
          prompt: "Summarise the following README:\n${{ steps.readme.outputs.content }}"

      - name: Print AI Summary
        run: echo "${{ steps.inference.outputs.response }}"

How it works

1. Checkout repository – Makes the README file available.
2. Read README content – Stores the file’s content in an output variable.
3. Call AI model – Sends the README text to the openai/gpt-4.1 model hosted in GitHub Models.
4. Print result – Displays the AI’s summary in the workflow log.

This provides a handy demonstration of AI inference directly inside CI/CD. Example print result output:

The Action Playground repository is a space for experimenting with GitHub Actions and CI/CD automation. It offers sample workflows, action templates, and integration examples with third-party services. Users can clone or fork the repo, explore or modify workflow files under , and trigger GitHub Actions by committing changes. Contributions are welcome via issues or pull requests, and the project is licensed under MIT.

Prototyping Made Easy: Playground

Before locking prompts into workflows, test them in the GitHub Models playground.

You can:

• Compare models side by side
• Tune parameters and refine prompt wording
• Share and iterate with your team

Pro tip: Move working prompts into .prompt.yml for version control, collaboration, and maintainability.

When Might You Use This Setup?

Developers are already using GitHub Models to streamline repetitive tasks:

1. Issue triage – auto-tag issues, summarise bug reports
2. Code review helper – summarise PRs, flag risky changes, draft release notes
3. Docs maintenance – refresh API or README snippets when code changes
4. Information extraction – surface recurring issues from comments
5. Test coverage – generate missing unit tests for new functions

The big win is saving developer time and embedding intelligence directly into your automation.

Reflections from Practice

From hands-on use, the biggest wins with GitHub Models are speed and predictability:

• Prototyping new automations takes minutes, thanks to the playground.
• Output validation prevents surprises in production pipelines.
• Workflows feel smoother because everything stays inside GitHub – no external APIs or secrets to manage.

The key is to start small: summarise PRs, generate release notes, or auto‑tag issues. From there, layer in complexity as confidence grows.

Wrapping Up: Smarter Workflows Right Now

GitHub Models are not just a shiny AI add‑on. They’re a practical toolkit for modern DevOps pipelines – saving time, reducing errors, and embedding intelligence into the workflows you already rely on.

Get started today:

1. Try the playground to explore models.
2. Move prompts into .prompt.yml for structure.
3. Use actions/ai-inference to wire them into workflows.
4. Validate responses with JSON schema for safety.

The result? Smarter automation. Faster reviews. Better developer productivity.

If you’ve been waiting for AI that just works with your existing GitHub setup – this is it.

Additional reading