GitHub Copilot Extensions are shutting down on November 10, 2025, and if your teams rely on @mentions inside Copilot Chat to call internal tools, you’ve hit the cliff. The good news: VS Code Copilot extensions remain supported. The better news: Model Context Protocol (MCP) is the replacement path—and you can stand up an MVP server in days, not weeks, if you focus on the right cuts. (github.blog)
What just changed—and when
Here’s the thing: GitHub is retiring server-side Copilot Extensions built as GitHub Apps in favor of MCP, a standard for connecting tools to AI assistants across vendors. Key dates to ground your plan:
• September 24, 2025: New Copilot Extension creation was blocked.
• November 3–7, 2025: Brownout testing period (temporary interruptions).
• November 10, 2025: Full sunset by 11:59 PM Pacific—extensions disabled in Copilot Chat.
• VS Code Copilot extensions are unaffected and remain supported.
• Hybrid GitHub Apps can stay in Marketplace if you disable the Copilot Extension configuration. (github.blog)
Does this affect my VS Code extension?
No—client-side VS Code Copilot extensions continue to work. This shutdown specifically targets server-side Copilot Extensions implemented as GitHub Apps. If your team says “we use a Copilot extension” and you’re unsure which kind, check your Marketplace listing or your app settings. (docs.github.com)
Why MCP—and what it means in practice
MCP is a vendor-neutral protocol for letting AI agents call tools and fetch context. Instead of writing one-off integrations for Copilot, then another for a different assistant, you build an MCP server once and host your tools behind it. That server exposes capabilities (think: commands, resources, prompts) that multiple hosts can consume—GitHub Copilot included. Practically, your “extension” becomes a service with clear schemas, predictable auth, and transport you can test locally. (github.blog)
GitHub Copilot Extensions: the 48‑hour MCP migration playbook
If you’re staring at red dashboards today, don’t try to rebuild everything. Ship a narrow, reliable MVP and iterate. This is the order we’ve used on client rescues.
1) Triage and freeze
Inventory all @mentions that previously mapped to Copilot Extensions. Rank them by business impact (blockers first, then high-frequency tasks). Freeze any new feature work touching those flows until your MCP server replaces the top three. Capture examples of real prompts and expected outputs—your test cases start here. (github.blog)
2) Pick one transport and one auth path
Keep the first iteration boring. Stand up your MCP server in the language your team debugs fastest. For auth, start with the least-privileged token or OIDC flow you can rotate. Assume you’ll refactor later for broader scopes or hardware-backed keys.
3) Model your tools as contracts
Turn each action into a concise tool schema: name, description, required/optional params, and safe defaults. Treat these like public APIs. Don’t pass raw file paths or internal IDs without validation; normalize inputs so your logs are explainable when things break.
4) Make outputs copy‑pastable and structured
Your first win is utility. Format responses with just enough structure that a developer can paste them into a PR or terminal without cleaning up. When in doubt, return JSON blocks with a human-readable summary.
5) Focus on three golden paths
Pick your top three flows—e.g., “create a release branch,” “page CI logs for a failed deploy,” “generate a Jira-ready bug summary.” Implement these end-to-end with ruthless guardrails and timeouts. You can add advanced features once you’re unblocked.
6) Test with transcripts, not just unit tests
Re-create real chat transcripts that triggered your extension. Run them through the MCP server using your host’s dev environment. Capture latency, token counts, and edge prompts. If a path takes more than 8–10 seconds, split it or cache aggressively.
7) Cutover with feature flags
Route only the top golden path to MCP for the first cohort of users. Keep a rollback switch for a week. Document the exact prompt patterns that your server supports so users don’t guess—and then complain.
Marketplace: can I keep my app listed?
Yes, if it’s a hybrid GitHub App. You must disable the Copilot Extension configuration in your app settings before the sunset to keep the listing active. Pure extension apps will be removed. Plan your listing text update to point users at your MCP-based path. (docs.github.com)
What about GitHub Actions—do I need it enabled?
Previously some Copilot features nudged you to turn on Actions. As of early November changes, the Copilot coding agent no longer requires GitHub Actions to be enabled at the org level. That’s one less dependency to juggle during your migration week. (github.blog)
MCP server blueprint you can copy
Here’s a minimal skeleton we’ve coached teams to implement under time pressure:
• Health endpoint returning build SHA, config version, and dependency checks.
• Auth middleware enforcing least-privileged scopes; instrument rotation metrics.
• Tool registry mapping names to handlers with schema validation and runtime guards.
• Stable, semantic error codes (MCP_HOST_TIMEOUT, TOOL_INPUT_INVALID, UPSTREAM_401).
• Structured logging (correlation ID, user/org, tool name, duration, token count).
• Caching layer for read-heavy tools; a simple LRU or Redis is plenty to start.
• Rate limiter per user and per tool to avoid blast radius.
Data points that matter this week
• Sunset is November 10, 2025, by 11:59 PM Pacific; after that, Copilot Extensions built as GitHub Apps stop working entirely in Copilot Chat.
• VS Code Copilot extensions remain supported.
• Brownouts already occurred November 3–7—mine your logs for early failures and prompts that need guardrails.
• Hybrid apps can remain in Marketplace if you disable the Extension configuration.
• If you rely on premium models in Copilot, a November 18 billing change removes default $0 budgets for many enterprise/team accounts—set policy to avoid unwanted charges. (github.blog)
“People also ask” quick hits
Can I convert my Copilot Extension into a standard GitHub App?
If your app has non‑Copilot functionality, those features continue. You’ll need to remove the Copilot Extension configuration and ship MCP for the chat-driven parts. (github.blog)
Is there any grace period?
No functional grace after the sunset. The brownouts were the last heads-up. Treat this as a hard stop and cut over now. (github.blog)
Will MCP lock me into GitHub?
No. MCP is positioned to work across multiple AI assistants. Designing tools once and reusing them elsewhere is the point. (github.blog)
Security and compliance: don’t skip these
Your MCP server is another surface area. Keep auditors happy and incidents rare by doing the basics well:
• Principle of least privilege on every upstream token; avoid broad org scopes.
• Secrets from a vault, not env files; rotate on deploys and when role changes happen.
• Input validation for tool params; reject unrecognized fields.
• PII minimization in logs; redact by default, allow explicit sampling in non‑prod.
• Backpressure with circuit breakers for flaky upstreams.
• Access logs with user and host identifiers to answer “who ran what” in minutes.
UX that prevents thrash
Document the exact prompts your MCP tools support and embed them in a “/help” tool. Add friendly errors (“Try: @tool create release branch from main with version=1.7.0”). Cache the last successful output for quick re-runs. Provide a “dry‑run” flag that prints commands without executing. These touches save engineering-hours immediately.
Governance and cost control (Nov 18 heads‑up)
Many enterprises created before August 22, 2025 still have a default $0 premium request budget in Copilot, which blocks extra premium requests. Beginning November 18, 2025, GitHub removes those default $0 budgets for enterprise and team accounts, shifting control to your premium request paid usage policy. Decide whether to enable overage or block it, and set per‑SKU budgets if you use coding agent or other AI add‑ons. Share this with finance today. (docs.github.com)
If you’d like a deeper dive on spend controls, our guide on avoiding Nov 18 Copilot surprise bills breaks down the knobs with practical defaults.
Rollout checklist you can run in an afternoon
• Disable the Copilot Extension configuration on any hybrid Marketplace app so the listing stays live.
• Stand up an MCP server with one transport and one auth path.
• Implement three highest‑impact tools and guard them with strict schemas.
• Add structured logging and a health endpoint; set SLOs (p99 under 8s).
• Capture real prompts from your brownout logs and replay them in staging.
• Gate to a pilot group; announce supported prompts and known limits.
• Plan week‑two: add two more tools, expand caching, tighten rate limits. (docs.github.com)
What to do next (developers)
• Ship the MVP MCP server today; don’t chase parity.
• Turn transcripts into tests; promote the three golden paths.
• Add a “/help” tool and dry‑run mode to reduce noisy failures.
• Set up alerting on tool error rates and latency.
What to do next (engineering leaders and owners)
• Update your Marketplace listing and docs within 24 hours.
• Decide premium request overage policy before November 18 to avoid disputes.
• Communicate what works now vs. what’s coming next week; keep status in a shared doc.
• Budget one engineer-day for security hardening and log scrubbing.
Need a second set of hands?
We’ve shipped these migrations under fire. If you want a focused assist, see our services or talk to us about a short, time‑boxed engagement. If you prefer to self‑serve, start with our MCP migration plan and, if you run Actions-heavy automation, skim what to change in GitHub Actions this month to keep your CI stable.
Zooming out
Sunsetting server-side Copilot Extensions is painful if you invested heavily in them—but MCP is the right long-term bet. Tooling built this way will meet your developers where they are, whether that’s Copilot, another IDE, or a different agent entirely. Move fast on an MVP, document the supported prompts, and keep your blast radius tight. By next week, you’ll be shipping on a steadier, more portable foundation.
