GitHub boosts Copilot with agents, new models, and MCP support

Used the new agent mode of VS Code’s Copilot. I’m impressed. I threw at  it what I thought was going to be a monumental task, and it scanned 4-5 different files, figured out how it was working, and made modifications in all those files to work exactly how I wanted. First time.

Agent Mode allows users to select from several powerful LLMs, including Anthropic’s Claude 3.5 and 3.7 Sonnet, Google’s Gemini 2.0 Flash, and OpenAI’s GPT-4o. GitHub reported that Agent Mode currently achieves a 56.0% pass rate on the SWE-bench Verified benchmark using Claude 3.7 Sonnet, with expectations for improvement as underlying models advance.

Alongside the Agent Mode rollout, GitHub announced the general availability of two other features:

• Code Review Agent: After a preview period where it was reportedly used by over one million developers, this agent assists in reviewing pull requests on GitHub.
• Next Edit Suggestions: This feature aims to streamline coding by suggesting subsequent edits, allowing developers to quickly accept changes and move forward—enabling them to “tab tab tab your way to coding glory.”

MCP: The ‘USB port for intelligence’

Central to the enhanced agent functionality for GitHub Copilot is the Model Context Protocol (MCP), now available in public preview. Dohmke described MCP as allowing developers to “equip Agent Mode with the context and capabilities it needs to help you, like a USB port for intelligence.”

MCP enables Copilot Agent Mode to interact with various tools within a developer’s engineering stack – from telemetry systems to infrastructure management tools. When a prompt is entered, the agent can consult available MCP tools to gather context (like database schemas) or perform actions (like querying a web service).

As an example, GitHub outlined a prompt: “Update my GitHub profile to include the title of the PR that was assigned to me yesterday.” The agent, leveraging MCP, would identify the necessary tools, interact with them iteratively (potentially calling multiple tools in sequence), and ultimately fulfil the request.

To foster this ecosystem, GitHub has released a new open-source and local GitHub MCP server. This server allows any LLM tool supporting MCP to integrate with GitHub functionalities like searching repositories, managing issues, and creating pull requests. A community inventory repository highlighting available MCP servers is also available.

GitHub introduces premium models and new Copilot pricing tiers

Expanding on its multi-model strategy, GitHub is making Anthropic Claude 3.5, 3.7 Sonnet, 3.7 Sonnet Thinking, Google Gemini 2.0 Flash, and OpenAI o3-mini generally available.

Access to these specific models will operate via a new ‘premium request’ system, complementing the unlimited access to the base model (currently GPT-4o) included in all paid Copilot plans for core features like Agent Mode, chat, and code completions.

Starting 5 May 2025, GitHub Copilot Pro users will receive 300 monthly premium requests. Copilot Business and Enterprise customers will receive 300 and 1000 monthly premium requests respectively, beginning between 12-19 May 2025. Usage of these premium models is unlimited until these dates.

A new tier, Copilot Pro+, has also been introduced for individual developers. Priced at $39 per month, it includes 1500 monthly premium requests and promises access to “the best models, like GPT-4.5.”

For users exceeding their monthly allowance, a pay-as-you-go option for additional premium requests is available which starts at $0.04 per request. Both individuals and organisations can opt-in and set spending limits. Different premium models will consume varying numbers of requests per use, allowing users to balance capability with cost.

“With GitHub Copilot, what started out as a developer platform company is [now] a platform where anyone can be a developer. Together, GitHub and Microsoft fully intend on enabling a world with 1 billion developers,” Dohmke concludes.