Copilot offline mode + BYOK

[pelikhan]

Since we gateway all LLMS, we should switch to BYOK route + offline mode.

https://github.blog/changelog/2026-04-07-copilot-cli-now-supports-byok-and-local-models/

[github-actions]

🤖 This is an automated response from Repo Assist.

Good feature request! It's worth clarifying the scope before planning implementation work.

What the MCP Gateway does today

gh-aw-mcpg is an MCP protocol proxy — it routes tool calls (GitHub, Slack, etc.) from an AI agent to backend MCP servers. It does not route LLM inference calls. LLM calls (to Copilot, OpenAI, etc.) happen at the orchestration layer above the gateway.

What BYOK/offline would mean here

The Copilot CLI changelog you linked is about the orchestration layer supporting local models and custom API keys for LLM inference. For the MCP Gateway, the relevant impact is:

1. Token/auth passthrough: Already partially supported — the gateway reads tokens via GITHUB_MCP_SERVER_TOKEN, GITHUB_TOKEN, GITHUB_PERSONAL_ACCESS_TOKEN, GH_TOKEN in priority order (see internal/envutil/github.go). A custom Copilot API key could be passed through via these env vars.

2. API endpoint override: The gateway already respects GITHUB_API_URL to redirect API calls to GHE or custom endpoints (internal/envutil/github.go:LookupGitHubAPIURL). This could cover routing to a local model endpoint in some configurations.

3. Offline/local model support: If "local model" means a locally-running MCP server (e.g. a local Ollama-backed MCP server), it can already be configured as an http backend in the gateway config. If it means routing LLM inference calls through the gateway, that's out of scope for the gateway and belongs in the orchestration layer.

Suggested path forward

• For BYOK with the Copilot API: check if setting GITHUB_API_URL + a custom API key token covers your use case
• For local LLM models as MCP tool providers: already works with type: "http" backend config
• For full LLM inference routing through the gateway: file an issue on the gh-aw (orchestration) repo, as that's where agent model selection lives

Happy to dig deeper into any specific part of this if you can share more detail about the intended deployment topology.

Generated by Repo Assist · ● 6.1M · ◷

To install this agentic workflow, run

gh aw add githubnext/agentics/workflows/repo-assist.md@851905c06e905bf362a9f6cc54f912e3df747d55