You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Describe the feature or problem you'd like to solve
It would be ideal if running the Copilot CLI in a VS Code project could load the MCP information already configured in VS Code.
Proposed solution
When you run the Copilot CLI in a VS Code project terminal, it retrieves and loads MCP information from .vscode/mcp.json in the project and from Code/User/mcp.json in the VS Code user settings.
This prevents duplication of MCP configurations and allows centralized management through integration with VS Code.
The same would go for custom prompts and all of the other work that goes into setting up VS Code Copilot Chat. If you have gone through all of that work to set up VS Code Copilot Chat for a given project, the copilot CLI should just use all of that.
If you can call local models with VS Code Copilot Chat, you should be able to use them with GitHub Copilot CLI automatically.
What is strategy here? What is the relationship between VS Code Copilot Chat and GitHub Copilot CLI? Are they completely separate products with no relation between? If GitHub Copilot CLI is just going to be a dumb independent AI agent that goes off of an AGETNS.md file like Codex CLI, then what is the advantage using GitHub Copilot CLI over Codex CLI? Codex CLI is open source and I can use it to call locally hosted models. And it works pretty well.
Without leveraging the setup of a VS Code project with VS Code Copilot Chat setup, why would anyone bother with GitHub Copilot CLI?
I notice org wide custom agents are now available in the latest vscode, being able to config the same available mcps would help complete the usefulness of shared custom agents across an org
Activity
bartlettroscoe commentedon Sep 26, 2025
The same would go for custom prompts and all of the other work that goes into setting up VS Code Copilot Chat. If you have gone through all of that work to set up VS Code Copilot Chat for a given project, the copilot CLI should just use all of that.
See:
bartlettroscoe commentedon Sep 26, 2025
Same goes for setup to call local models with VS Code Copilot Chat:
(to be released for real in the Sept version released in October?) and the yet-to-be-completed:
If you can call local models with VS Code Copilot Chat, you should be able to use them with GitHub Copilot CLI automatically.
What is strategy here? What is the relationship between VS Code Copilot Chat and GitHub Copilot CLI? Are they completely separate products with no relation between? If GitHub Copilot CLI is just going to be a dumb independent AI agent that goes off of an AGETNS.md file like Codex CLI, then what is the advantage using GitHub Copilot CLI over Codex CLI? Codex CLI is open source and I can use it to call locally hosted models. And it works pretty well.
Without leveraging the setup of a VS Code project with VS Code Copilot Chat setup, why would anyone bother with GitHub Copilot CLI?
Update:
See:
showpune commentedon Nov 6, 2025
Will Coding agent share the same configuration from vscode?
dnguyen-wf commentedon Dec 11, 2025
I notice org wide custom agents are now available in the latest vscode, being able to config the same available mcps would help complete the usefulness of shared custom agents across an org
RyanHecht commentedon Feb 25, 2026
We now load MCP settings from
.vscode/mcp.json!felici1b commentedon Mar 27, 2026
Now we can get copilot cli mcp settings directly from vscode mcp settings? How? Could be great not avoid settings duplication.
sajalverma17 commentedon May 28, 2026
The same thing is being said here as well: #54 (comment)
And yet, this documentation page says Copilot CLI in fact, does NOT load MCP from
.vscode/mcp.json.I'm on v1.0.54, and this is really confusing. @RyanHecht Can you please clarify?