GA versions sometimes just remove the "Preview" label. This one adds features aimed at enterprise networks, self-hosted inference servers and developers who juggle more than one account.
From an August preview to a GA release
BYOM's preview started fairly recently. Visual Studio Magazine reported that Microsoft announced the new capability in an Aug. 24 Visual Studio Blog post, making BYOM available in preview in Visual Studio 18.10 Insiders. Microsoft's preview post said BYOM was available and enabled by default across Community, Professional, and Enterprise SKUs.
The preview was limited. Microsoft's C++ team blog said it supports Microsoft Foundry, OpenAI, Anthropic, and Ollama, including custom endpoints for OpenAI and Ollama. The redesign also broke compatibility with the older setup. Daily.dev's summary noted that previously added BYOM models under the old Ask/Agent modes are no longer supported and must be re-added.
The 18.12 release brings four main changes:
- Multiple configurations for the same provider, each with its own name
- A dedicated Custom Endpoint provider for OpenAI-compatible services and local models
- A choice of API format and settings for model capabilities
- Better proxy support, so provider connections and Agent Mode follow Visual Studio's resolved proxy configuration
Admins can also turn BYOM on or off through organizational policy.
Section summary: In about three months BYOM went from an Insiders preview with four providers to a GA feature with a generic endpoint option and enterprise controls.
No GitHub sign-in required (with conditions)
Microsoft says BYOM doesn't need a GitHub sign-in. That matters to teams that can't reach GitHub or aren't allowed to. Microsoft says you can still use your chosen models, including local or internally hosted models in air-gapped environments, as long as the model endpoint is reachable and you meet its authentication requirements.
That condition is important. The announcement doesn't promise that all of Visual Studio, or every Agent Mode tool, works offline. What it does say is that GitHub is no longer a required step between you and a model endpoint you can already reach. If your inference server sits behind a firewall your workstation can't get through, or your API key has expired, BYOM can't help.
Section summary: BYOM removes GitHub as a dependency, but you still need network access and valid credentials for the endpoint itself.
Separate "Work" and "Personal" configurations
During the preview, using two accounts with one provider was awkward. In 18.12 you can keep several configurations for the same provider and give each one a clear name, such as "Work" and "Personal." Each keeps its own connection details and its own model list.
Microsoft gives these examples:
- A personal account for side projects alongside a company account
- Separate Microsoft Foundry resources
- An experimental deployment kept next to your team's approved endpoint
You no longer have to delete and re-add a provider every time you switch.
Custom Endpoint: gateways, OpenRouter and LM Studio
Many teams don't call a model vendor's public API directly. They go through a self-hosted inference server, an AI gateway or an endpoint their organization provides. The new Custom Endpoint provider gives these OpenAI-compatible services their own place in the provider settings. Microsoft names OpenRouter as an example.
Setup takes two fields: the endpoint URL, plus an API key if the endpoint needs one. If the endpoint returns a compatible model list, Visual Studio can show those models. Endpoints that don't return a list won't populate automatically.
Microsoft also mentions local models served by LM Studio. You start LM Studio's local server, then add its OpenAI-compatible URL as a Custom Endpoint to use the model in Agent Mode. The post doesn't give a default URL or port, so check LM Studio's own server settings for the address.
Be realistic about local models. In its hands-on look at the preview, Visual Studio Magazine found that a local Qwen3:4b model running through Ollama connected easily and handled a basic project-analysis query, but became confused and failed to complete a simple Agent coding task. Microsoft's C++ team gives a similar warning: model capabilities vary, and not every model supports every Agent Mode tool. A 4-billion-parameter model on a laptop GPU won't perform like a frontier model in the cloud.
Section summary: Custom Endpoint opens BYOM to any OpenAI-compatible service, but being able to connect a model doesn't mean it can handle agent tasks.
Matching the API format and capabilities to your model
"OpenAI-compatible" covers a lot of different implementations, and Microsoft now lets you say which API format your endpoint uses. When you add a custom model, you can choose:
| API format | Typical fit |
|---|---|
| OpenAI-compatible Completions | Older or simpler OpenAI-style servers |
| OpenAI-compatible Responses | Endpoints that implement OpenAI's newer Responses-style API |
| Anthropic-compatible Messages | Services that expose Anthropic's Messages format |
(The "typical fit" column is general industry context, not Microsoft guidance.)
You can view and change the configured format later in Manage models. You can also set context size, maximum output, tool calling and vision support for each model.
The announcement doesn't say Visual Studio tests these capabilities automatically. Enter what your model and deployment actually support. If you turn on tool calling for a model that can't do it, expect Agent Mode to behave like the confused Qwen model above.
Enterprise controls: proxy and policy
Microsoft says it improved proxy support, so provider connections and Agent Mode both use Visual Studio's resolved proxy configuration. In corporate networks where every outbound request goes through an inspecting proxy, that consistency can decide whether BYOM works at all.
Admins can enable or disable BYOM through organizational policy. The GA post doesn't name the policy or explain how to deploy it. In August, though, Microsoft's preview post said it planned an ADMX policy to disable it for Professional and Enterprise users. That suggests a standard Group Policy or Intune setting. Confirm the exact setting in Microsoft's administrative documentation before you roll it out.
Security teams should also think about governance. Microsoft's documentation for the older BYOM experience warned that when you use a custom model, output is returned directly from the provider and might bypass the filtering for responsible AI in Copilot. The GA announcement doesn't say whether that still applies. The safe assumption is that the model's provider, not Microsoft or GitHub, is responsible for data handling and content filtering. That's exactly why admins have a disable switch.
Section summary: Proxy support works consistently now and admins have a policy switch, but data handling depends on whichever provider you connect.
How to set up BYOM in Visual Studio 18.12
Microsoft's steps:
- Install Visual Studio 18.12 GA.
- Open Chat and choose Manage models from the model picker.
- Select Add model provider.
- Choose Microsoft Foundry, OpenAI, Anthropic, Ollama or Custom Endpoint.
- Configure the connection, give it a clear name and add the models you want.
- Select your model and start working in Agent Mode.
When it works, your model appears in the Chat model picker and responds in Agent Mode.
If something goes wrong
- No "Add model provider" button? Microsoft says your organization's admin may have disabled BYOM. Ask IT before assuming it's a bug.
- Custom Endpoint shows no models? Your endpoint may not return a compatible model list. Add the model manually.
- Errors or garbled replies? Check that the API format matches what your server actually uses (Completions, Responses or Messages).
- Agent tasks fail? Check the tool-calling and context-size settings, and consider whether the model is capable enough for agent work.
- Timeouts on a corporate network? Check Visual Studio's proxy setup, since BYOM now uses it.
- Models from the old preview are missing? Models added under the old Ask/Agent experience have to be re-added.
The bigger picture
Microsoft is betting that developers want to choose their own models, and that enterprises want that choice to work within their existing controls. Air-gapped support, gateway compatibility and the policy switch cover the main checks a security team would ask about.
What's still missing? The GA post has no supported-model list, no BYOM-specific pricing or licensing details, and no explanation of how each provider handles data. Those are fair questions for regulated organizations. For a developer who just wants Visual Studio to talk to an existing Foundry deployment or a home LM Studio setup, 18.12 makes that much easier.
Microsoft is asking for feedback through the Developer Community forum and a short survey. If you test BYOM with an unusual gateway or a local model, share how it went in the forum here.