Tools API for your Copilot Extensions

[isidorn]

We have finalized the Tools API 🚀
Docs can be found here https://code.visualstudio.com/api/extension-guides/tools

Try out the API and let us know what you think about it.

Check out this sample:
https://github.com/microsoft/vscode-extension-samples/blob/main/chat-sample/README.md

Add the following to package.json:
"enabledApiProposals": ["languageModelToolsForAgent”],

In langaugeModelTools in package.json, ensure you add the vscode_editing tag:

You then should be able to go to agent mode and your tool can be invoked from a user prompt:

This should be enough to get you started and show what’s possible.

We are still trying to understand what gaps we might have in this approach and what adjustments/affordances we may need to have to make the experience work. As you start building out a POC with some developer scenarios, please report back if there’s anything you feel is required to land your testing scenarios.

Thank you 🙏

[isidorn]

We have launched agent mode in preview https://code.visualstudio.com/docs/copilot/copilot-edits#_use-agent-mode-preview
And in the next couple of months we will explore Agent calling tools contributed via the language model tools api.

[doggy8088]

Will it be possible in the future to invoke any VS Code Commands through Agent Mode?

[isidorn]

No.

[lemmy]

@isidorn, could you clarify why VSCode's agent mode won't support invoking existing VSCode commands, even if it's just a whitelisted subset? As an extension developer, it seems redundant to have to reimplement each command using vscode.LanguageModelTool, especially when those commands are already defined and available in my extension.

[isidorn]

Because language model needs nice description of functionality - to decide if it should call it or not. commands do not have that by default.

So please do contribute tools! Thanks 🙏

[lemmy]

Here's my follow-up at #2411 (comment) ;-)

[ArtemGovorov]

@isidorn Hey, coming here from the https://code.visualstudio.com/blogs/2025/02/24/introducing-copilot-agent-mode post:

Exploring tool extensibility and MCP servers as tools for agent mode. If you have extension ideas tell us

We have recently added support for Copilot chat in Wallaby, also working on adding support in Console Ninja. We have implemented a few vscode.LanguageModelTool tools for the chat. Our tools allow LLM to retrieve code execution paths and runtime values for better runtime errors investigation. It would be great to be able to expose those tools to Copilot Edits and Agent as well.

In terms of tool contribution, some form of tool registration for the Edits/Agent from within a running extension would be great, because it would be much easier to get access to the full context of the extension's state from the tool. The current way to expose tools to chat via a chat participant @... mention seems to be not very flexible neither discoverable for the potential variety of tools that may be contributed.

MCP server support would also work for us in some way, but I guess extensions may need more plumbing to be able to connect the MCP server to the running extension itself. With MCP, it would be great to at least have MCP roots support, so that the tool can know about the context it is running in, should it somehow need to connect/talk to a running extension to get some state or invoke some code of the extension.

[vnijs]

Just saw this video from the LangChain folks. MCP connections are possible in Cursor, Windsurf, and Claude Desktop. I like Copilot Edits and plan to use it with my students. Adding the option to connect MCP to VS Code Copilot would be very very nice. Thanks for your consideration.

[isidorn]

Thanks - we are adding MCP tool support to Copilot Agent mode. You can already try it out with VS Code Insiders. And should be in VS Code stable in the next couple of weeks.

[mohamedmansour]

Hello! I work on the Chromium code base which is massive, and I Would like to write an agent that uses git grep to search for code / files / etc, how can I make that happen? git grep is fast to search and I would like it to take my prompt and do a bunch of git grep to figure out which one it should include in my context (to fill the context window)

[isidorn]

Write an extension, that contributes a git grep tool - then agent should use the tool your extension is contributing.

You can extend agent mode today in VS Code Insiders with the LanguageModelTools API. Check out this sample:
https://github.com/microsoft/vscode-extension-samples/blob/main/chat-sample/README.md

Add the following to package.json:
"enabledApiProposals": ["languageModelToolsForAgent”],

In langaugeModelTools in package.json, ensure you add the vscode_editing tag:

You then should be able to go to agent mode and your tool can be invoked from a user prompt:

This should be enough to get you started and show what’s possible.

We are still trying to understand what gaps we might have in this approach and what adjustments/affordances we may need to have to make the experience work. As you start building out a POC with some developer scenarios, please report back if there’s anything you feel is required to land your testing scenarios.

[ropevp]

What tool, maybe an MCP tool would you recommend using like obsidian, to link into the codebase, to draw on ? is there a local rag available for this any plans to extend the brain of copilot with local knowledge - maybe a better way to go about.
MCP Marketplace like Cline that would get people excited - but it must get EVERYTHING and amazing support.

[aymuos15]

Hello! Thank you so much for this!

How do I allow the agent to use the terminal and run stuff (or force it to run only one command if possible) without asking for my permission each time?

[isidorn]

Not possible this milestone. We are adding it in April.

[aymuos15]

Thanks for the response! Eagerly waiting :)

[isidorn]

fyi @Tyriar

[gokul427]

Apart from tools, do we get other primitives like resources and prompts for agent mode mcp?

[isidorn]

This is on the plan but has not happened yet.

fyi @digitarald

[lemmy]

I've integrated a few languageModelTools into our extension, and the models can now successfully invoke them. However, the modelDescription field is proving too limited and cumbersome for providing the level of guidance needed. We've found that, to ensure the LLMs use these tools correctly, we also need to include more comprehensive instructions—specifically, through a .github/copilot-instructions.md file.
What is unclear is how to include or bundle this copilot-instructions.md file as part of our VSCode extension. Right now, the file lives in the user's workspace, which is outside the control of our extension. Is there a way to package it with the extension so we can manage and distribute it reliably, or are we missing something and there is a better alternative to copilot-instructions.md.

[isidorn]

We would love if the LLM can use the tools correctly without extensions modifying the copilot-instructions.md file.
Did you try fine-tuning your tool description to make sure the LLM uses it correctly? Are you using Sonnet or 4.1, since we notice 4o is not good at tool calling.

[lemmy]

As extension developers, we don't always know which LLMs our users will be working with.

Right now, the modelDescription field in package.json is too limited. Could you introduce something like a modelDescriptionHook instead? This hook could be called whenever an LLM requests a tool description, i.e., what's modelDescription today, allowing extension developers to generate a tool description dynamically. This would make it possible to tailor the description based on the specific LLM in use and even adjust it according to the tool’s capabilities.

[isidorn]

This is a fair feature request - feel free to open one https://github.com/microsoft/vscode/issues

[lemmy]

microsoft/vscode#247426

[lemmy]

A VSCode extension exposes an underlying (deterministic) tool, such as a compiler, within the VSCode. One approach to integrating this functionality into an agentic workflow is through languageModelTools. A different architecture uses MCP to expose the compiler. Given that both languageModelTools and MCP aim to expose tools to LLMs, what is the long-term vision for each? Are they going to coexist, diverge, or converge? Currently, MCP offers the advantage of being independent of VSCode, making it easier to test agentic workflows. This is particularly relevant because testing agentic workflows is not yet fully available within the VSCode + languageModelTools. On the other hand, languageModelTools offers a much more user-friendly experience. Users can access tools without needing to run separate MCP servers, making it more convenient for those working entirely within VSCode.