Skip to main content

VS Code · GitHub Copilot · Local harness

VS Code is retiring the Copilot Local harness: what it means for your local models

Bodega One8 min read
Share
Quick answer

VS Code 1.146 deprecates the Copilot Local harness, and GitHub says it will be removed in February 2027. Despite the name, “Local” never meant local models: it is the older code that runs agent sessions inside the editor window. Its replacement, the Copilot harness, has been the default for new sessions since VS Code 1.143. The part that matters if you use Ollama or your own API keys: today, VS Code's docs say bring-your-own-key local models work in the Local harness with no GitHub account, no Copilot plan and no internet. On the Copilot harness, your own models still need an experimental setting, chat.agentHost.byokModels.enabled. Test your setup on the new harness now, not in February.

On November 16, GitHub announced that the Local harness is on its way out. The announcement is short, and it does not mention local models or bring-your-own-key (BYOK) setups at all. The answers are spread across several pages of the VS Code docs. This post puts them in one place: what is changing, what happens to Ollama and BYOK models, and a checklist to work through before February.

What is the Local harness, and what replaces it?

A harness is the part of VS Code that runs an agent session: it ties together the model, the context and the tools. VS Code can run sessions on several. The harness docs list four: Copilot, Local (deprecated), Claude and Codex. Cloud is an execution target for the Copilot harness, not a separate harness. Picking a model does not change the harness.

The Local harness runs in the VS Code extension host and is tied to the editor window, so the window has to stay open while the agent works. The Copilot harness runs in a separate process called the Agent Host, is built on the Copilot SDK, and shares its runtime with Copilot CLI and the GitHub Copilot app. That is why a Copilot harness session can move between the editor and the Agents window. GitHub says choosing it does not move your files to the cloud.

The name trips people up. VS Code's own docs say the Local name “does not mean that the model runs locally or that the session works offline.” Model requests in a Local session can still go to a hosted service. Retiring it does not, by itself, take local models away.

When does it happen?

  • VS Code 1.143: the Copilot harness became the default for new sessions.
  • VS Code 1.146: the Local harness is deprecated. New features go to the Copilot harness only.
  • Transition: for three months, the Local harness stays available where your organization's policy allows it.
  • February 2027: the Local harness is removed, along with the temporary enterprise setting that keeps it as the default. GitHub has not given a day.

How do the two harnesses compare?

Local harness (deprecated)Copilot harness
Where it runsVS Code extension host, tied to the editor windowAgent Host process, separate from the editor window
Your own modelsModels set up in VS Code: Copilot, or your own credentialsGitHub Copilot, or supported models through experimental Agent Host BYOK
Agents windowNot available; editor onlyAvailable
ToolsRuns in the editor window, alongside your extensionsNot every built-in or extension tool; extension tools need an open editor window
HooksYour current hooksNot converted automatically; hooks follow the selected harness, so review them by hand
Managed permissions and sandboxCannot enforce themCan enforce them, but selecting the harness does not turn sandboxing on
New featuresNo new features from 1.146Copilot Fusion, assisted permissions, Copilot connectors (preview), canvases (preview)

What happens to Ollama, LM Studio and BYOK models?

This is the question the announcement skips. Here is what the VS Code language models docs say today.

  • On the Local harness: BYOK models work without a GitHub account or a Copilot plan, and BYOK local models, for example through Ollama, work fully offline.
  • On the Copilot harness: supported BYOK models need the experimental setting chat.agentHost.byokModels.enabled (“BYOK models in Agent Host”). The docs do not list which models count as supported.
  • Signed out: the Agents window docs say BYOK models can run in the desktop Agents window without GitHub sign-in once the provider credentials are set. The browser-based Agents window always requires GitHub sign-in.
  • Ollama: VS Code's built-in Ollama provider is deprecated. The docs point you to the official Ollama extension from the Ollama publisher instead.
  • LM Studio, llama.cpp, vLLM and other servers: the old OpenAI Compatible provider is deprecated too, replaced by Custom Endpoint, which takes any compatible endpoint using the Chat Completions, Responses or Messages API. LM Studio is not named on that page, so test yours.
  • Tool calling: a model has to support tool calling to be used with agents in chat. Models without it are hidden from the picker.
  • What still needs GitHub: BYOK does not cover features that depend on the Copilot service. Semantic search, inline code completions and embeddings need a GitHub account and an internet connection, so they do not work offline. Titles and commit messages need chat.utilityModel and chat.utilitySmallModel pointed at one of your own models when you are signed out.

So local models are not being switched off, and nothing in the docs says they will be. But the only path documented to work fully offline is the one being removed, and its replacement depends on an experimental setting. Nobody outside Microsoft knows what that setting will look like in February. That is the risk worth planning around.

What about Copilot's own on-device model?

Separately from BYOK, VS Code has an experimental on-device model that Copilot manages. It is a different thing with different rules. It needs GitHub sign-in with Copilot access, runs only in Copilot harness sessions, and needs an NVIDIA GPU with 96 GB or more of VRAM (unified memory counts) plus about 55 GB of disk. Business and Enterprise users also need an admin to enable the Editor Preview Features policy. The current model is MAI Code 1.1 Flash. You install llama.cpp yourself (build b11658 or later), set chat.copilot.localModel.llamaServerPath to the full path of llama-server, and turn on chat.copilot.localModel.enabled. Its responses use no AI credits, and chat.copilot.localModel.useInAuto lets Auto route to it with cloud fallback. VS Code's docs also warn that local models may not have the same safety filters as cloud models and may be easier to steer with malicious instructions, especially with auto-approved actions.

What should you do before February?

VS Code's transition guide lays out the move. Here it is in order, with the local-model checks added.

  1. Update VS Code to your organization's approved release. The guide assumes you are signed in to GitHub with Copilot access; if you run your own models signed out, do your testing in the desktop Agents window.
  2. Pick one small task that works today on the Local harness. Write down the result you expect, the tools it uses, your customizations and the approvals it asks for.
  3. Back up your customization files and settings. Original and migrated files do not sync, so keep one active copy of each.
  4. If you use your own models, turn on chat.agentHost.byokModels.enabled, check the model appears in the picker, and confirm it supports tool calling. For Ollama, install the official Ollama extension. For other servers, set them up as a Custom Endpoint.
  5. Start a new session in the Chat view or the Agents window and set Session Target to Copilot before your first prompt. Then choose your model and agent.
  6. Run the migration: Chat: Open Customizations → Migrations lists what VS Code found. Migrate does the supported conversions. Migrate with Agent is experimental and uses Copilot credits.
  7. Fix what does not convert:
    • Prompt files become skills, but their agent, model and tool choices are not kept. Set them again.
    • Hooks are not converted. Check their events, script input and output, and failure behavior by hand.
    • MCP servers: ${input:...} and ${command:...} prompts and secrets stored in the editor do not transfer. Re-enter them. Do not paste a literal credential in place of a placeholder.
    • Memory is not converted. Local memory and Copilot Memory are separate.
    • Tool selection is now saved to your profile, under Chat: Open Customizations → Tools, not per request.
  8. To carry on an existing conversation, switch that chat's Session Target to Copilot. The handoff is one way.
  9. Run your test task again and check the actual edits and test results, not just that it finished. Run /sandbox policy in a Copilot session to see the file system and network rules that apply.
  10. If something blocks you, report it as a VS Code issue now, while the Local harness still exists as a fallback.

For admins

  • To require the Copilot harness, set temporaryAllowLocalHarness and temporaryDefaultToLocalHarness to false in enterprise-managed settings, not user or workspace settings.
  • To encourage it without requiring it, set the ChatEditorPreferCopilotHarness device policy to true.
  • Review any vscode.temporaryDefaultToLocalHarness exception. It is removed with the Local harness, so leaving it set will not keep Local alive.
  • Sandboxing on the Copilot harness is chat.agent.sandbox.enabled. A sandbox device policy for the Local harness is not a Copilot control.
  • Business and Enterprise users need the Bring Your Own Language Model Key in VS Code policy enabled on GitHub.com before they can add their own keys.
  • Claude and Codex harnesses are off under enterprise-managed settings unless an admin allows them.
  • Run Developer: Policy Diagnostics to confirm which values are in effect.

Want an editor where local models are the default?

If you run models on your own machine, you can also skip the question of which harness supports them. In Bodega One Code, local models are the starting point, not a setting to turn on. The first launch asks how you want to run it: On this computer, With a cloud key, or Offline only (air-gap). While you decide, it scans this machine for running Ollama, LM Studio, vLLM and llama.cpp servers and reads your GPU, and nothing leaves the machine until you press Continue.

  • Your models: 15+ provider presets, including local ones for Ollama, LM Studio, vLLM, llama.cpp, MLX on Apple Silicon, Jan, SGLang and others, plus a custom OpenAI-compatible endpoint. Local models need no API key.
  • Offline when you want it: air-gap mode is one switch that blocks the agent's outbound connections, web tools and network shell commands, with a badge in the top bar while it is on.
  • Your Copilot rules come with you: open a project with a .github/copilot-instructions.md and Bodega offers a one-time Import into Bodega that appends it to .bodega-rules. It reads AGENTS.md and CLAUDE.md on every message too.
  • Your editor preferences: the VS Code / Cursor settings import is under Settings → About.
  • Your MCP servers: add them under Settings → Connections → MCP servers, including hosted servers that sign in with OAuth.

It is a full desktop IDE with a Monaco editor, AI chat and a coding agent, and it is free to use, commercial work included. If you are setting up a local model for the first time, the local LLM guide covers which models fit your GPU, and the models and providers docs walk through connecting one.

Common questions

What is the Local harness in VS Code?
It is the older code that runs GitHub Copilot agent sessions inside the VS Code editor window, tying together the model, context and tools. Despite the name, it does not mean the model runs locally or that the session works offline.
When is the VS Code Local harness removed?
VS Code 1.146 deprecates it, and GitHub says it will be removed in February 2027 after a three-month transition. The Copilot harness has been the default for new sessions since VS Code 1.143.
Can I still use Ollama and local models in VS Code after February 2027?
Yes, nothing in the docs says local models are blocked. On the Copilot harness, your own models need the experimental setting chat.agentHost.byokModels.enabled. VS Code now points Ollama users to the official Ollama extension, and other servers to a Custom Endpoint. The model must support tool calling.
Do BYOK models in VS Code need a GitHub account?
VS Code docs say BYOK models work without a GitHub account on the Local harness, and can run signed out in the desktop Agents window once provider credentials are set. The browser-based Agents window always needs GitHub sign-in.
What do I need to change when moving to the Copilot harness?
Set Session Target to Copilot, then run Chat: Open Customizations and Migrations. Prompt files become skills, hooks are not converted, MCP input prompts and stored secrets do not transfer, and memory is not converted.

Written by the Bodega One team. We build Bodega One Code, the local-first AI IDE, and we write here about local models, AI costs, and what we learn shipping it. More about the team and why we build local-first on the about page.

Keep learning

Free, vendor-neutral courses and guides in the Bodega One AI Academy, from what a model is to shipping an app.

Stay in the loop

Build-in-public updates, model picks, and Copilot/Cursor news as it breaks.

Ready to own your tools?

Free to use, and the beta is open to everyone. Download free.