GitHub Copilot adds BYOK token caps, model toggles and MCP flows
The July 27 update hands solo builders granular control: per-model token caps, MCP and custom agents in the CLI, OTel traces, and a /rubber-duck command.
Copy markdown
BYOK finally gets hard token caps
JetBrains users can now set default `maxInputToken` and `maxOutputToken` limits for BYOK and custom endpoints — a direct lever on context bloat and per-call cost when you're paying for your own keys. You can also disable or enable all built-in Copilot models from the model-management controls.
MCP servers and custom agents land in the CLI
Copilot CLI can now use MCP servers and custom agents directly inside Claude agent flows, so your terminal agent reaches your own tools and sub-agents — not just Copilot's defaults. Sessions also gained fork support and a `/rubber-duck` command for talking a problem through.
Agent runs get OpenTelemetry export
Under Settings > Tools > GitHub Copilot > Chat you can now configure OpenTelemetry export for agent workflows, piping traces into whatever observability stack you already run. Debugging a flaky agent run stops being a black box.