Burak Unuvar

Posted on Aug 16

FireConnect Open Models from MS Foundry to GH Copilot

#githubcopilot #llm #fireworks #tutorial

❓ 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

DEV Community