GitHub Copilot agent mode and MCP support roll out to all VS Code users

The line between suggesting and acting

Code assistants in editors have mostly been used in two ways: a chat panel that answers questions, and edits applied to files the developer already has open. GitHub draws a further line at action. Unlike chat or multi-file edits, agent mode takes action to translate an idea into code ^[1]. The distinction sounds small and is not: a suggestion waits to be accepted, while an action changes the state of the project, so files get created, commands run and errors surface.

What shipped on 4 April 2025

On 4 April 2025 GitHub rolled out Copilot agent mode in Visual Studio Code to all users, together with MCP support ^[1]. The two did not arrive in the same state: MCP support in VS Code shipped as a public preview, while agent mode was progressively rolling out to the stable channel ^[1]. The release shipped with VS Code v1.99, which also brought general availability for Copilot code review and next edit suggestions ^[1].

How the agent works

With simple prompts, agent mode completes all the necessary subtasks across files it automatically identifies or generates ^[1]. The developer does not have to point at each file in advance; deciding which files matter is part of the job being handed over.

It can suggest terminal commands or tool calls and ask the user to run them ^[1]. It also analyzes runtime errors with self-healing capabilities ^[1]. Read together, these describe a loop rather than a single answer: take the request, work out which files are involved, write or edit them, propose commands, look at what comes back, and adjust. Where commands are concerned, the person at the keyboard still approves and runs them, so the agent proposes and the developer disposes.

Models and one benchmark number

Agent mode was powered by a choice of Claude 3.5 and 3.7 Sonnet, Gemini 2.0 Flash and GPT-4o ^[1]. GitHub reports that agent mode reached a 56.0% pass rate on SWE-bench Verified with Claude 3.7 Sonnet ^[1]. A pass rate on a benchmark of software engineering tasks measures how often an agent finishes a task on its own; it does not measure whether the resulting code matches a particular team's conventions, review rules or unwritten habits.

MCP as a port

GitHub describes MCP as a USB port for intelligence, letting agent mode use tools such as reading a database schema or querying the web ^[1]. The metaphor is worth taking literally: a USB port does not itself do anything, it standardizes how many different devices can be plugged in. MCP plays that role for an agent's tools, so capabilities can be added without rebuilding the assistant each time.

Alongside editor support, GitHub released an open-source, locally run GitHub MCP Server that adds cross-repo code search, issue management and pull request creation to any MCP-capable tool ^[1]. Because it runs locally and is open source, it is something a developer starts on their own machine rather than a service they must request access to, and it can be inspected before use.

Quotas and plans

Access was also a question of quotas. From 5 May 2025, Pro users received 300 premium requests a month and Enterprise users 1,000, while a new Pro+ plan at $39 a month offered 1,500 ^[1]. Agent work is metered differently from autocomplete because each task can involve many model calls.

Why it matters

The release changed the default unit of work in the editor from a line of code to a task. A developer describes an outcome, and the tool decides which files to open, which commands to propose and how to react when something fails. The second change is that the assistant was tied to an open protocol rather than a fixed set of integrations, so tools built outside GitHub can be plugged into it.

Neither change removes the developer from the loop: commands are still proposed for a person to run, code still arrives as edits to be read. What shifts is where the effort goes, from typing the steps to stating the goal and checking the result.