Copilot for JetBrains gets MCP toggles and enterprise default-model controls

What it is

GitHub Copilot's JetBrains plugin picked up two changes that matter if you're building or evaluating agent loops: a setting to control when MCP servers actually start, and admin-managed defaults for which Copilot model opens a new chat. A new MCP setting lets you disable automatic server startup for Copilot and Claude, giving you more control over when configured tools become active. On the model side, enterprise administrators can now use managed settings to choose any available Copilot agent model as the default for new conversations, which gives organizations more control over the starting model while still preserving an explicit model picker choice.

The next-work-session test

You're debugging an agent that calls three MCP servers — your repo's Postgres tool, a scratch filesystem server, and a flaky internal search API. Today you probably restart the IDE and wait for everything to spin up before you can repro. With the new toggle, you keep auto-start off, launch only the Postgres server for this session, and run the same prompt against two models from the picker to see which one picks the right tool first. That is a real eval loop, in the IDE, without touching a terminal.

The other win is for anyone shipping an internal "golden" setup: admins can pin a default model org-wide while leaving the picker alive for one-off experiments.

Pricing

Pricing: unverified for this update specifically. The changelog does not call out a plan gate for the MCP toggle, and the enterprise-managed default-model control is clearly scoped to Copilot Enterprise administrators. Public per-seat pricing for Copilot Free, Pro, Business and Enterprise lives on GitHub's plans page — check the current numbers directly at github.com/features/copilot before quoting them to finance. We couldn't verify current list prices in time for this piece and won't guess.

What we'd actually use it for

Three things, honestly:

Limits

• The MCP control is a startup toggle, not a per-tool allow/deny list. If you want fine-grained tool gating inside a running server, you still manage that in the server itself.
• Enterprise-managed defaults require Copilot Enterprise admin rights. Solo devs and small teams on Pro/Business can switch models from the picker but can't centrally pin one.
• "Default model" means default for new conversations — not a routing rule. Nothing here auto-escalates from a cheap model to a stronger one mid-session.
• This is a JetBrains-plugin release. Behavior and settings surfaces in VS Code are a separate codebase and may lag or lead.
• No eval tooling is bundled. You're still writing your own prompt harness, logging, and scoring if you want real numbers on tool-use accuracy.

Try it if

• You run agent experiments inside IntelliJ, PyCharm, GoLand, or Rider and the "every MCP server boots on launch" behavior has been getting in your way.
• You administer Copilot Enterprise and want to standardize the starting model without taking the picker away from developers.
• You want the inline Fix path from a diagnostic to a chat-driven patch to shorten triage.

Skip it if

• You're not on JetBrains — wait for the equivalent VS Code changelog.
• You need per-tool permissioning or audit logs across MCP calls. This update doesn't deliver that.
• You already drive MCP from a terminal-based agent (Claude Code, custom harness). The IDE toggle is nice but doesn't change your eval pipeline.
• You were hoping for a built-in model router or automatic fallback. Not here.

Full change notes on the GitHub Changelog entry for Copilot for JetBrains.