GitHub Copilot's agent mode reaches stable VS Code, with MCP tool support
GitHub is rolling out agent mode in Visual Studio Code to all users, bringing with it Model Context Protocol (MCP) support for connecting the coding assistant to external tools and services. The company also announced a new open-source GitHub MCP server and expanded premium model availability across paid Copilot tiers.
Agent mode differs from standard chat or multi-file edits by taking initiative beyond proposing changes. It breaks down a high-level request into subtasks across automatically identified or generated files, and can suggest terminal commands or tool calls for the user to approve. It also checks run-time errors and can recover from them without human intervention. While the feature has been tested in VS Code Insiders since February, it's now progressively reaching stable builds, with full availability expected within weeks; users can enable it manually in the interim.

In testing, agent mode with Claude 3.7 Sonnet achieved a 56.0% pass rate on SWE-bench Verified. The mode supports multiple models, including Claude 3.5 and 3.7 Sonnet, Google Gemini 2.0 Flash, and OpenAI GPT-4o.
MCP support in public preview
MCP support lets developers give agent mode access to the wider engineering stack—databases, telemetry, infrastructure, and external APIs—through a standardized interface GitHub likens to "a USB port for intelligence." Inside a VS Code chat prompt, the model can choose from available MCP tools to gather context or execute actions, iterating until the task is complete.
GitHub points to a growing community directory of MCP servers, and has released a local GitHub MCP server that gives agent mode capabilities such as repository and code search, issue management, and PR creation. Setup instructions for local and remote MCP servers are available from VS Code, and the GitHub MCP server repository documents native VS Code support.
Premium model requests and new tier
Claude 3.5, Claude 3.7 Sonnet (including Thinking), Gemini 2.0 Flash, and OpenAI o3-mini are now generally available through premium requests, which are metered separately from the unlimited use of the base model (currently GPT-4o) included in paid plans.
Starting May 5, 2025, Copilot Pro users receive 300 premium requests monthly. Copilot Business and Enterprise users receive 300 and 1000 per month respectively, beginning between May 12 and May 19, 2025. Each premium model draws a specific number of requests from this pool.
A new Pro+ plan targets individual developers with 1500 monthly premium requests and access to top models such as GPT-4.5, priced at $39 per month. Paid Copilot users can also opt into pay-as-you-go overages starting at $0.04 per request, with spending limits available for Business and Enterprise administrators.
CA note on agentic coding adoption in development workflows
The broader push toward agentic tooling comes as Tech Report's own analysis shows a steady increase in Copilot usage across enterprise codebases since the introduction of agent capabilities.



