DEV Community: Burak Unuvar

FireConnect Open Models from MS Foundry to GH Copilot

Burak Unuvar — Sun, 16 Aug 2026 21:37:19 +0000

❓ Impressed by the latest open-source model benchmarks and eager to use those models in GitHub Copilot?

❓ Using several AI coding tools and tired of configuring bring-your-own-key
(BYOK) access separately in each one?

🔥 FireConnect is an open-source CLI that connects agentic coding
tools to open-source model families such as Llama, Qwen, DeepSeek, Kimi, and
GLM.

💡 A reminder about native GH Copilot models

At the time of writing, GitHub Copilot already supports 30 models, including
open models such as Kimi K3.

FireConnect becomes useful when you want:

Faster access to an expanding range of models.

One configuration pattern across coding tools such as GitHub Copilot, Claude Code, and Cursor.

FireConnect puts it simply:
“Install once, sign in once, then flip any supported harness on or off > without hand-editing config files.”

What is Fireworks and FireConnect?

💡 Fireworks AI is a high-performance inference and fine-tuning platform focused on open-source models such as Llama, Qwen, DeepSeek, Kimi, and GLM.

💡 Microsoft announced Fireworks AI as an official first-party inference
provider inside Foundry in March 2026.

💡 FireConnect is the bridge between your AI coding harness and the model ecosystem, letting you plug in models your tool does not natively offer without changing how you work.

🤔 Wait, aren't open models already available on Foundry?

Yes, Microsoft Foundry already offers a rich catalog of models, including open source models sold directly by Azure. Fireworks provides an alternative inference path for many of those models at a quicker pace. Fireworks brings its own rapidly evolving catalog backed by an optimized inference platform that's directly integrated into Foundry.

Some examples:

Model Direct from Azure Via Fireworks on Foundry

Kimi K3 ❌ No ✅ Yes

GLM 5.2 ❌ No ✅ Yes

Kimi K2.7 Code ✅ Yes ✅ Yes

DeepSeek V4 Pro ✅ Yes ✅ Yes

Model	Direct from Azure	Via Fireworks on Foundry
Kimi K3	❌ No	✅ Yes
GLM 5.2	❌ No	✅ Yes
Kimi K2.7 Code	✅ Yes	✅ Yes
DeepSeek V4 Pro	✅ Yes	✅ Yes

🤔 This looks like yet another way of BYOK, do I really need FireConnect ?

If you only need a custom model for GH Copilot CLI or GH Copilot inside VS Code, BYOK is more than enough and works great !

FireConnect becomes much more interesting when you use multiple AI coding harnesses such as GitHub Copilot, Cursor, Claude Code**, Codex, or others and want a more consistent way to connect and route them to models available through Fireworks AI. Instead of configuring every harness independently against different model-provider interfaces, FireConnect provides a common integration layer across supported tools.

🔐 When to Use Which

When to Use Which

Use-Case Built-in providers and Extensions Custom Endpoint with FireConnect

Use a single model in VS Code chat ✅ Just add your key and go Works, but overkill for this

Use the same Fireworks models across Copilot, Cline, Continue, etc. ❌ Set up each tool separately ✅ One setup — toggle harnesses on/off

Get Fireworks latency and routing out of the box ❌ Generic endpoint, you're on your own ✅ Built for Fireworks, tuned for their infrastructure

Get your whole team on the same model config ❌ Everyone sets it up manually ✅ Shared config that anyone can reproduce

Use-Case	Built-in providers and Extensions	Custom Endpoint with FireConnect
Use a single model in VS Code chat	✅ Just add your key and go	Works, but overkill for this
Use the same Fireworks models across Copilot, Cline, Continue, etc.	❌ Set up each tool separately	✅ One setup — toggle harnesses on/off
Get Fireworks latency and routing out of the box	❌ Generic endpoint, you're on your own	✅ Built for Fireworks, tuned for their infrastructure
Get your whole team on the same model config	❌ Everyone sets it up manually	✅ Shared config that anyone can reproduce

** Warning**
FireConnect provides technical integration between Claude Code and Fireworks models through an Anthropic-compatible endpoint. But, this is not an Anthropic-endorsed configuration, and Anthropic does not provide support for routing Claude Code to non-Claude models through third-party gateways.
So maybe as a fun fact we can say:
Claude Code can technically be configured to go outside the Claude family to make most out of open weighted models, but Anthropic currently doesn't like this 😁

How to Get Started

Prerequisites

An active Azure subscription with the Fireworks integration enabled. See the Fireworks on Foundry setup guide.
A Microsoft Foundry project and a deployment of your preferred open-weight model from the Fireworks model catalog.

Step 1: Set Up the Foundry Resource and Project

Before running the Azure CLI commands, create a local .env file in the repository root. These variables are used by the commands below.

RESOURCE_GROUP=my-fireconnect-rg
AZURE_REGION=eastus
FOUNDRY_RESOURCE_NAME=my-fireconnect-foundry
PROJECT_NAME=my-fireconnect-project
MODEL_NAME=FW-GLM-5.2
BASE_URL=https://<your-resource-name>.services.ai.azure.com/
AZURE_API_KEY=your_key

# Load the variables into your current Bash session.
set -a
source .env
set +a

# create the resource group
az group create \
  --name "$RESOURCE_GROUP" \
  --location "$AZURE_REGION"

# Create the Microsoft Foundry resource.
az cognitiveservices account create \
  --name "$FOUNDRY_RESOURCE_NAME" \
  --resource-group "$RESOURCE_GROUP" \
  --location "$AZURE_REGION" \
  --kind AIServices \
  --sku S0 \
  --custom-domain "$FOUNDRY_RESOURCE_NAME" \
  --yes

# Create the Microsoft Foundry project.
az cognitiveservices account project create \
  --project-name "$PROJECT_NAME" \
  --name "$FOUNDRY_RESOURCE_NAME" \
  --resource-group "$RESOURCE_GROUP" \
  --location "$AZURE_REGION"

Step 2: Deploy a Fireworks Model on Foundry

Model benchmarks evolve frequently. For this example, I’ll use GLM-5.2 from Z.ai, which ranked among the strongest open-weight coding models in recent evaluations. See the GLM-5.2 benchmark analysis.

# Inspect the model format.
MODEL_NAME="FW-GLM-5.2"

MODEL_INFO=$(az cognitiveservices account list-models \
  --name "$FOUNDRY_RESOURCE_NAME" \
  --resource-group "$RESOURCE_GROUP" \
  | jq --arg MODEL "$MODEL_NAME" '.[] | select(.name == $MODEL)')

echo "$MODEL_INFO"

# Deploy the model.
az cognitiveservices account deployment create \
  --resource-group "$RESOURCE_GROUP" \
  --name "$FOUNDRY_RESOURCE_NAME" \
  --deployment-name "$MODEL_NAME" \
  --model-format "Fireworks" \
  --model-name "$MODEL_NAME" \
  --model-version "1" \
  --sku-name "DataZoneStandard" \
  --sku-capacity 100

Step 3: Install, Upgrade, and Configure FireConnect

Install FireConnect, then configure its global VS Code settings:

# Install FireConnect.
curl -fsSL https://raw.githubusercontent.com/fw-ai/fireconnect/main/install.sh | bash
fireconnect upgrade

# Configure the global FireConnect settings.
fireconnect configure \
  --provider azure \
  --base-url "$BASE_URL" \
  --api-key "$AZURE_API_KEY"

fireconnect vscode on --model FW-GLM-5.2
fireconnect vscode status

Start or restart VS Code, open Copilot Chat, and pick a Fireworks model from the model picker.

For one-off routing without changing the global configuration, pass your Foundry model with --model (for example, FW-GLM-5.2)

fireconnect vscode on \
  --azure \
  --base-url "$BASE_URL" \
  --model FW-GLM-5.2

To disable or uninstall FireConnect:

# Disable FireConnect for VS Code.
fireconnect vscode off

# Uninstall FireConnect for VS Code.
fireconnect uninstall

🔐 BYOK in Depth

Bring your own key (BYOK) allows you to use Copilot with models of your choice: local or hosted by an external provider. GitHub Copilot supports BYOK at two levels:

Feature	Enterprise BYOK	Local BYOK
Who configures it	Enterprise / Org owner	Individual developer
Where keys live	Server-side (GitHub)	Client-side only
Copilot license needed?	Yes	No
Works offline / air-gapped?	❌ No	✅ Yes

Local BYOK works across multiple surfaces:

VS Code	JetBrains IDEs	Xcode	Copilot CLI	GitHub Copilot App	Copilot SDK
✅ GA	✅ Public preview	✅ Public preview	✅ Supported	✅ Supported	✅ Supported

BYOK to GH Copilot in VS Code and FireConnect

VS Code offers different ways to add language models:

Provider option	What it means
Built-in providers	The provider is already listed (Azure, Anthropic, Gemini, OpenAI, and others)
Extensions	A marketplace extension provides the model (e.g., AI Toolkit for local models)
Custom endpoint	You have a self-hosted or enterprise endpoint that speaks Chat Completions, Responses, or Messages API

💡 FireConnect is configured through custom endpoints.

🔐 References

References

GitHub Agentic Workflows for Guardrailed Automation

Burak Unuvar — Sat, 08 Aug 2026 22:27:11 +0000

❓ Have you found your self orchestrating multiple coding agents across different harnesses, screens, surfaces? How many at most can you track and monitor simultaneously? 😫

❓ Did you ever give up on approving permission asks and started running all of them on YOLO mode ? Please try not to do that, at least not in production, agents can easily get out control and dangerous! 😱

Although I love all the magic and efficiency AI Coding Tools are bringing to my life, I'm a bit overwhelmed recently, trying to achieve tens maybe hundreds of tasks! Some really deserve my full attention and active steering to achieve the best possible outcome, while for most, the judgement of a model is more than enough.

In order to bring intelligent and controlled automation to my Github Repositories, I hired octopus students 🐙 🐙 🐙

GitHub Agentic Workflows(GH-AW).

What is GH-AW?

Disclaimer: GitHub Agentic Workflows a.k.a. GH-AW is an open-source GitHub CLI extension developed by GitHub. It is announced in February 2026 by GitHub Next and Microsoft Research, still in public preview but already widely adopted by Microsoft itself as well as many other organizations.

GH-AW enables running the coding agents you know and love such as GitHub Copilot, Claude Code, OpenAI Codex, or Google Gemini in GitHub Actions. Consider them as intelligent systems embedded directly into your CI/CD 💡 They all run within pre-defined guardrails, using safe outputs and sandboxed execution.

Some of the use cases are :

Repository Maintenance with Issue Triages, PRs Reviews
Continuous Documentation maintenance and consistency
Metrics & Analytics for daily reports and workflow health monitoring
Test improvements, CI failure diagnosis, quality checks

Hmm, what does it really address ?

👍 Do you want to gather daily activity data, analyze all the PRs and issues, check workflow results, and creates status report in form of issues ? Enable Daily Repo Status

👍 Do you want to automatically assign issues labeled cookie to the Copilot coding agent every 30 minutes? Enable Issue Monster.

Above are only two examples from a sample family of reusable pre-baked workflows maintained by Github Next, called agentics. There're plenty of interesting ones such as CI Doctor to Monitor CI workflows or Grumpy Reviewer for On-demand opinionated code review by a grumpy but thorough senior developer.

These may be useful, what else ?

👍 Do you have an automation in mind that you could describe entirely in natural language? Then you can create custom agentic workflows for that !

Whether it’s fixing a simple bug, creating and running tests on a schedule, fetching data from the web to keep your documentation up to date, or something entirely different.

How does it work ?

Agentic Workflow: From Prompt to Execution

                                         gh aw compile
┌──────────────────────┐  ┌──────────────────────┐  ┌──────────────────────┐
│  1. YOUR PROMPT      │  │  2. GH-AW            │  │  3. GH-AW COMPILER   │
│                      │  │   CUSTOMAGENT        │  │                      │
│  /agentic-workflows  │  │  Creates workflow.md │  │  Validates source    │
│  "Create a workflow" │──►  using its skill     │──►  Injects runtime     │
│                      │  │                      │  │  Generates YAML      │
│                      │  │  - schedule          │  │                      │
│                      │  │  - permissions       │  │  workflow.lock.yml   │
│                      │  │  - safe outputs      │  │                      │
│                      │  │  - task brief        │  │                      │
└──────────────────────┘  └──────────────────────┘  └───────────┬──────────┘
                                                                │
           ┌───────────────────── commit + push ────────────────┘
           ▼
┌──────────────────────┐                     ┌──────────────────────┐
│  4. GITHUB ACTIONS   │                     │  5. GITHUB REPO      │
│                      │                     │                      │
│  Launches agent      │                     │  Issues / PRs / code │
│  Executes task       │   ───────►          │                      │
│  Posts safe outputs  │                     │                      │
│                      │                     │                      │
└──────────────────────┘                     └──────────────────────┘

Why is this really a thing , can't I just build my own custom agents ?

Yes of course you can - sky is the limit ! But enforcing what they can do access and affect can can quite trickly. GH-AW runs within an isolated sandbox with Read-Only permissions, prevents Prompt Injection and is authorized to write only what you let.

When to use which: Github Actions VS Github Agentic Workflows

⚠️ A better question would be "How to use both together". 💡

Agentic workflows are complementary; your deterministic build, test, and release pipelines stay unchanged. Github Actions requires YAML configuration for fully deterministic systems. Agentic workflows use AI to interpret natural language instructions in markdown. Think of it as Continuous AI alongside CI/CD: a new automation layer in GitHub Actions for tasks where exact reproducibility doesn’t matter.

How can I start ?

Quick Start with Github CLI
A github repo to experiment some pre-built and 2 custom workflows
Disclaimer: Please don't hate me in case of errors 😃 and please feel free to submit issues or contribute 🤗
Are you looking something for advanced, a zero to hero style workshop ? gh-aw-workshop

Last but not least, here comes the main source of all information and updates GitHub Agentic Workflows 🚀 🚀 🚀