GitHub Copilot adds agent mode, MCP support in latest release

A brand new model of GitHub Copilot has simply begun rolling out to customers, including plenty of new capabilities, such because the addition of agent mode to VS Code. 

With agent mode, Copilot can iterate throughout a whole venture, suggesting terminal instructions, analyzing run-time errors, and extra. “With easy prompts, agent mode takes Copilot past answering a query, as an alternative finishing all mandatory subtasks throughout robotically recognized or generated information to make sure your main purpose is achieved,” the corporate’s CEO Thomas Dohmke wrote in a weblog submit. 

For instance, GitHub printed a demo video during which a developer must replace an internet site for runners that enables them to kind races by title, distance, and time. Upon receiving the request, Copilot analyzes the location to find out what must be modified, then begins by updating the backend and the UI, then producing unit checks for each, earlier than handing it again over to the consumer to run these checks. 

Agent mode can make the most of Claude 3.5 and three.7 Sonnet, Google Gemini 2.0 Flash, or OpenAI GPT-4o. It was first launched to VS Code Insiders in February, and is now starting to roll out to all VS Code customers, and the rollout ought to be full in just a few weeks.

The corporate additionally introduced a public preview for Mannequin Content material Protocol (MCP) help in agent mode. MCP is an open commonplace for connecting knowledge sources to AI instruments, and including it to agent mode will allow Copilot to make use of totally different instruments to deal with totally different duties, equivalent to understanding database schema or querying the online. 

For instance, asking Copilot to “replace my GitHub profile to incorporate the title of the PR that was assigned to me yesterday” would immediate the agent to take that request and the checklist of all accessible MCP instruments and ask an LLM what it ought to do subsequent based mostly on that data. 

“The GitHub native MCP server equips agent mode with compelling capabilities equivalent to looking throughout repositories and code, managing points and creating PRs – turning agent mode into a robust consumer of the GitHub platform,” Dohmke wrote. 

And eventually, after including help for different AI fashions like Anthropic’s Claude 3.5 Sonnet and Google’s Gemini 1.5 Professional, GitHub is now saying that requests to fashions aside from Copilot’s base mannequin (GPT-4o) might be thought-about premium requests and topic to limits, after having been limitless so far. 

All of GitHub’s paid plans will nonetheless supply limitless requests for agent mode, context-driven chat, and code completions for the bottom mannequin. Beginning in Might, Copilot Professional and Copilot Enterprise prospects will obtain 300 month-to-month premium requests and Copilot Enterprise prospects will obtain 1000 month-to-month premium requests. The corporate can be including a brand new Professional+ plan that offers 1500 month-to-month premium requests. 

If a buyer needs to transcend their month-to-month allotment, there’s a pay-as-you-go construction, beginning at $0.04 per further premium request.