Virtual tools are, in GitHub’s description for GitHub Copilot Chat in VS Code, a way of functionally grouping similar tools under one “virtual tool” that the chat agent can expand as needed — GitHub likens them to directories containing related tools. The model is given a general sense of what is available without being flooded with hundreds of tool names, and because similar tools tend to be used and activated together, GitHub says grouping them also reduces the cache miss rate it would expect if the model searched for tools individually.
Usage
GitHub introduced the approach, as reported in How we’re making GitHub Copilot smarter with fewer tools, as a middle ground between exposing every tool — the default VS Code toolset has about 40 built-in tools and Model Context Protocol servers can push the total into the hundreds — and filtering the toolset in a way that restricts what the agent can do.
The groups are formed by what GitHub calls adaptive tool clustering. An earlier attempt fed all tools to an LLM and asked it to group and summarize them, which GitHub abandoned because the number of groups could not be controlled and sometimes exceeded model limits, the process was slow and costly in tokens, and the model sometimes failed to categorize some tools. Instead, embeddings for each tool are generated with GitHub’s internal Copilot embedding model and grouped by cosine similarity, which GitHub says gives precise, stable and reproducible groups. A model call still writes a summary of each cluster, and tool embeddings and group summaries are cached locally so recomputing them is cheap.
The same idea is applied to Copilot’s own built-in tools: a core set of 13 tools is shown up front, and the remaining built-in tools are grouped into four virtual categories — Jupyter Notebook Tools, Web Interaction Tools, VS Code Workspace Tools and Testing Tools — that the model expands only if necessary.
When It Applies
Virtual tools address the case where an agent has more tools than it can reason over efficiently — GitHub reports that MCP servers can bring in enough tools to exceed some models’ API limits. Grouping alone leaves a failure mode: the model may open several wrong groups before finding the right one, each costing a cache miss, an extra round trip and a chance of failure. GitHub pairs virtual tools with Embedding-guided tool routing for that reason. The technique and its reported results come from a single vendor describing its own product.