Copilot CLI BYOK and Local Model Support¶

Connect Copilot CLI to your own model provider — Ollama, Azure OpenAI, Anthropic, or any OpenAI-compatible endpoint — for cost control, data residency compliance, and fully air-gapped workflows.

Released April 7, 2026, Copilot CLI BYOK lets you replace GitHub-hosted model routing with your own provider (GitHub Changelog). Configuration is four environment variables.

Configuration¶

Set these variables before launching copilot (GitHub Docs):

Variable	Required	Description
`COPILOT_PROVIDER_BASE_URL`	Yes	Base URL of your provider's API endpoint
`COPILOT_MODEL`	Yes	Model identifier (or use `--model` flag)
`COPILOT_PROVIDER_TYPE`	No	`openai` (default), `azure`, or `anthropic`
`COPILOT_PROVIDER_API_KEY`	No	API key — omit for unauthenticated local endpoints

Supported Providers¶

OpenAI-compatible (Ollama, vLLM, Foundry Local, OpenAI):

export COPILOT_PROVIDER_BASE_URL=http://localhost:11434
export COPILOT_MODEL=llama3.2

Azure OpenAI:

export COPILOT_PROVIDER_BASE_URL=https://YOUR-RESOURCE.openai.azure.com/openai/deployments/YOUR-DEPLOYMENT
export COPILOT_PROVIDER_TYPE=azure
export COPILOT_PROVIDER_API_KEY=YOUR-AZURE-API-KEY
export COPILOT_MODEL=YOUR-DEPLOYMENT-NAME

Anthropic:

export COPILOT_PROVIDER_TYPE=anthropic
export COPILOT_PROVIDER_BASE_URL=https://api.anthropic.com
export COPILOT_PROVIDER_API_KEY=YOUR-ANTHROPIC-API-KEY
export COPILOT_MODEL=claude-opus-4-5

Offline Mode¶

COPILOT_OFFLINE=true disables telemetry and restricts the CLI to communicate only with the configured provider (GitHub Changelog). Combined with a local Ollama instance, this enables fully air-gapped workflows.

Important caveat: isolation is only complete when COPILOT_PROVIDER_BASE_URL also points to a local or on-prem endpoint. A remote URL sends prompts and code context over the network to that provider regardless of offline mode (GitHub Docs).

GitHub Authentication¶

GitHub login is optional when using BYOK. With only provider credentials, the full local agentic experience runs without a Copilot subscription. Adding GitHub authentication re-enables /delegate, GitHub Code Search, and the GitHub MCP server (GitHub Changelog).

Model Requirements¶

Any model must support tool calling and streaming. A 128k context window is recommended for complex tasks. All built-in sub-agents (explore, task, code-review) inherit the provider configuration automatically — there is no per-agent routing (GitHub Changelog).

Trade-offs¶

Consideration	Detail
Capability ceiling	Local models have lower reasoning quality than hosted frontier models for complex tasks
Single provider	Only one provider config is active at a time — no native multi-provider routing per task
Failure handling	Invalid config surfaces actionable errors; the CLI never silently falls back to GitHub-hosted models
Discovery	`copilot help providers` prints quick setup instructions in-terminal

Example¶

Air-gapped compliance scenario: a financial services team must keep all code on-premises.

# Pull a capable local model
ollama pull llama3.2

# Configure Copilot CLI for local-only operation
export COPILOT_PROVIDER_BASE_URL=http://localhost:11434
export COPILOT_MODEL=llama3.2
export COPILOT_OFFLINE=true

# Launch — no GitHub auth required, no external network calls
copilot

The team loses access to /delegate and GitHub MCP server but gains full data-residency compliance. For tasks requiring stronger reasoning, swapping COPILOT_MODEL and COPILOT_PROVIDER_BASE_URL to a hosted provider (without COPILOT_OFFLINE) routes those sessions externally while preserving the same CLI experience.

Key Takeaways¶

Four env vars configure any OpenAI-compatible provider, Azure OpenAI, or Anthropic
COPILOT_OFFLINE=true restricts network to the configured provider — full isolation requires a local provider URL
GitHub auth is optional; adding it re-enables cloud features like /delegate and GitHub MCP
All built-in sub-agents inherit provider config; per-task routing requires managing env vars externally
Invalid provider config surfaces clear errors — no silent fallback to GitHub-hosted models