GitHub added Codex as an agent provider in the GitHub Copilot plugin for JetBrains IDEs, putting OpenAI’s coding agent behind the same in-IDE agent picker developers use for Copilot Chat.
The July 7 update is in public preview. GitHub says users install the Codex CLI locally, enable Codex in JetBrains settings, set the CLI path, and then select Codex from the Copilot Chat agent picker. For Copilot Business and Enterprise subscribers, an administrator must enable the editor preview features policy before the feature can be used.
The update is bigger than a provider toggle. GitHub also expanded agent customizations with hooks support, richer MCP server management, generated customization files, Copilot CLI approval settings, Claude agent permission modes, Claude debug logs, model view management, and custom model support configured by Business and Enterprise administrators.
The IDE becomes the agent router
Developers are no longer choosing only a completion model. They are choosing an agent runtime and permission model.
GitHub’s JetBrains plugin now lets a user run Codex, Copilot CLI sessions, and Claude agent sessions from the same broad agent surface. That means IDE policy has to cover which provider can run, which CLI path is trusted, which tools are available, which permissions are allowed, and how activity is debugged.
The MCP piece is especially practical. GitHub says MCP servers can be managed directly in Agent Customizations for Copilot CLI sessions, including command and HTTP server types. Workspace-level MCP servers can be defined with .github/mcp.json.
That turns repository configuration into agent infrastructure. A project can carry tool access instructions alongside code, but that also means teams need review rules for MCP definitions the same way they review build scripts, CI permissions, and dependency changes.
Approval modes are becoming product features
GitHub’s new Copilot CLI approval settings show how much agent behavior is moving into explicit UI.
Default Approvals follow the user’s configured policy and prompt when needed. Bypass Approvals auto-approve tool calls while still asking clarifying questions when necessary. Autopilot, currently in preview, auto-approves tool calls and auto-responds to clarifying questions so the agent can keep iterating.
Those modes are useful because long coding tasks often fail when the agent stops for routine confirmations. They are also risky because fewer interruptions means more reliance on the initial prompt, the agent’s judgment, and the surrounding sandbox.
GitHub’s adjacent repository-overview feature points in the same direction from a lighter angle. Copilot can now summarize a repository’s purpose, technologies, and contribution guidelines for all Copilot plans. The low-friction path from understanding a repo to running an agent inside it is getting shorter.
