Agentic Workflows with self-hosted runners and local inference

This repository demonstrates how GitHub Agentic Workflows can be used with:

• self-hosted runners
• model-routing platforms
• local inference

You can even run your Actions jobs, including AI models, directly on your laptop or Mac Mini.

We demonstrate the following example scenarios:

• Azure VM self-hosted runner calling OpenRouter for inference
• Local Mac self-hosted runner truly local inference via Qwen running with Ollama

Of course, you mix and match any Actions runner with any inference host, as long as they can communicate with each other.

Why self-host?

Self-hosting runners gives you extra control over your Actions execution environment, and your choice of hosting platform. Choosing a model-routing platform like Open Router can give you access to additional models, and hosting your inference yourself can help control costs. Frontier models still require datacenter-scale resources to host, but there are models small enough to run on a MacBook Air that can still perform useful work.

Prerequisites

Install the GitHub CLI and the Agentic Workflows extension:

gh auth login
gh extension install github/gh-aw

Compile the Markdown workflow sources into GitHub Actions lock files:

gh aw compile --validate --actionlint
scripts/patch-local-qwen-awf-pricing.sh

The generated .lock.yml files are committed because GitHub Actions runs those files, not the Markdown workflow sources.

Scenario 1: Azure VM Runner, OpenRouter Inference

Use this lane when you want the agent job to run on a Linux VM you operate, while model calls go through OpenRouter. We demonstrate using Azure, but you can easily choose any suitable VM host.

Read the full guide:

open infra/azure-vm/README.md

Short version:

gh aw secrets set OPENROUTER_API_KEY --value "$OPENROUTER_API_KEY"
infra/azure-vm/create-runner-vm.sh
gh workflow run azure-runner-capability-smoke.yml
gh aw run azure-vm-openrouter

Scenario 2: Local Mac Runner, Local Qwen/Ollama Inference

True local inference! We use Qwen 2.5 0.5B here for demonstration purposes, which is capable of some simple tasks. We also had success using Qwen 3.6 27B, a very capable model, on a MacBook Air (though it makes the MacBook nearly unusable for other work while running). Some models that are quite close to frontier performance can be run on memory-maxed Mac Studios.

Full guide:

open infra/local-macrunner-qwenollama/README.md

Quick test:

scripts/run-local-macrunner-qwenollama.sh "Check the local agent lane."

The launcher starts or verifies Ollama on macOS, creates Qwen model aliases, boots an x86_64 Lima Linux VM, registers that VM as the GitHub Actions self-hosted runner, dispatches the workflow, and watches it.

Runner Requirements

gh-aw self-hosted runners must be Linux hosts with:

• Docker.
• Passwordless sudo for the runner service account.
• iptables support.
• Outbound HTTPS access to GitHub, GHCR, and the selected engine endpoint.
• Access to any domains listed in the workflow network allowlist.

A macOS host cannot directly satisfy the Linux runner requirements for the gh-aw agent job, which is why the all-local Mac scenario uses Lima to run a local Linux VM. The Mac still owns the local model endpoint.

Adapting The Pattern

Use the included scenarios as starting points:

• Change the runner labels to point at another host class.
• Change the bootstrap scripts to install your internal dependencies.
• Change OPENAI_BASE_URL and the engine config to point at another OpenAI-compatible gateway.
• Change the model IDs and smoke tests to match your local inference service.
• Add private network routes, mounted caches, GPUs, or internal tools to the self-hosted runner.