GitHub Copilot CLI

GitHub Copilot CLI is GitHub's terminal coding agent, started with the copilot command. Its bring-your-own-key (BYOK) mode replaces GitHub-hosted models with a provider you choose. With Tokens it uses the default openai provider type, which sends OpenAI Chat Completions requests to https://tokens.bd/v1.

This page is about the terminal agent. For Copilot Chat inside VS Code, see GitHub Copilot in VS Code. The two are configured differently:

What you need#

• A Tokens key. Create one for Copilot CLI alone, with a monthly spend cap (API keys).
• The model id from /models. The examples use deepseek/deepseek-v4.1-flash.
• Copilot CLI installed. GitHub's documented options are npm install -g @github/copilot (Node.js 22 or newer), winget install GitHub.Copilot on Windows, brew install --cask copilot-cli, or curl -fsSL https://gh.io/copilot-install | bash on macOS and Linux. On Windows, GitHub requires PowerShell 6 or newer.
• A model that supports tool calling and streaming. Copilot CLI returns an error if either is missing. GitHub recommends a context window of 128K tokens or more.

Set it up#

Copilot CLI turns BYOK on when COPILOT_PROVIDER_BASE_URL is set. Set the base URL, your key and the model, then start copilot:

export TOKENS_API_KEY="tok_live_your_key"
export COPILOT_PROVIDER_BASE_URL="https://tokens.bd/v1"
export COPILOT_PROVIDER_API_KEY="$TOKENS_API_KEY"
export COPILOT_MODEL="deepseek/deepseek-v4.1-flash"
copilot

$env:TOKENS_API_KEY = "tok_live_your_key"
$env:COPILOT_PROVIDER_BASE_URL = "https://tokens.bd/v1"
$env:COPILOT_PROVIDER_API_KEY = $env:TOKENS_API_KEY
$env:COPILOT_MODEL = "deepseek/deepseek-v4.1-flash"
copilot

The PowerShell lines last for the current window. Keep the key in your shell profile or a secrets manager, not in a file inside a repository.

The key is the only credential Tokens needs. If you would rather use COPILOT_PROVIDER_BEARER_TOKEN, Tokens also accepts Authorization: Bearer, but the API key variable is the simpler choice.

Token limits#

Copilot CLI looks up context limits from a well-known model name. Tokens ids are not well-known names to it, so set the limits yourself:

export COPILOT_PROVIDER_MAX_PROMPT_TOKENS=120000
export COPILOT_PROVIDER_MAX_OUTPUT_TOKENS=8192

The numbers are placeholders. Copy the real context window and maximum output from the model's page in /models, and leave room for the output inside the context window. COPILOT_PROVIDER_MODEL_ID and COPILOT_PROVIDER_WIRE_MODEL exist for cases where the name the CLI should look up differs from the name sent to the provider. You do not need either with Tokens, because the id you set in COPILOT_MODEL is the id Tokens expects.

The wire API setting#

GitHub's page also lists COPILOT_PROVIDER_WIRE_API, but it does not say which values are accepted. Leave it unset. Tokens supports both Chat Completions and the Responses API, so the default works. Other guides set it to responses for specific models; if you try that, test with the check below first.

Check that it works#

Test the key and model outside Copilot CLI first:

curl https://tokens.bd/v1/chat/completions \
  -H "Authorization: Bearer $TOKENS_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model": "deepseek/deepseek-v4.1-flash", "max_tokens": 16, "messages": [{"role": "user", "content": "Reply with OK"}]}'

Then run Copilot CLI with one prompt:

copilot -p "Reply with OK"

You should get a short answer, and the request should appear in usage analytics on your dashboard. To check tool calling, start copilot in a project and ask it to read a file. If the model answers but never touches your files, it probably does not support tool calling well; see Tool calling.

Choosing a model#

Pick a model with reliable tool calling and a long context. Choosing a model compares them, and the /models catalog shows context and output limits for each. Switch for a session with copilot --model <id> or by changing COPILOT_MODEL. Only one provider configuration is active at a time.

Limits and what to know#

• Spend. An agent loops: one prompt can become many requests, each carrying the growing conversation. Use a key with a monthly spend cap (API keys) and watch usage on the dashboard.
• GitHub sign-in. GitHub's announcement says signing in to GitHub is optional in BYOK mode, and that signing in adds GitHub features such as /delegate, GitHub code search and the GitHub MCP server. GitHub's install page lists an active Copilot subscription as a prerequisite. The BYOK page does not say which plans support BYOK. Ask whoever manages your Copilot plan if you are unsure.
• Organization policy. If your organization or enterprise has disabled Copilot CLI, you cannot use it, with or without BYOK.
• What still goes to GitHub. GitHub's page does not list it. It does say COPILOT_OFFLINE=true stops Copilot CLI contacting GitHub's servers. That setting does not make Tokens local: prompts and code context still go to https://tokens.bd/v1.
• Model ids with a slash. Tokens ids look like provider/model. GitHub's page does not discuss ids with a slash. If COPILOT_MODEL is rejected, test the id with the curl above first.
• Sessions. Start copilot in a shell where the variables are set. A session started elsewhere does not see them.

Troubleshooting#

Copilot CLI still uses GitHub models. COPILOT_PROVIDER_BASE_URL is not set in the shell that started copilot. Run echo $COPILOT_PROVIDER_BASE_URL (or $env:COPILOT_PROVIDER_BASE_URL in PowerShell).

401 missing_api_key or invalid_api_key. COPILOT_PROVIDER_API_KEY is empty, or it holds another provider's key. A Tokens key starts with tok_live_. See Errors.

404 model_not_found. COPILOT_MODEL is not an exact Tokens id. Copy it from /models or GET https://tokens.bd/v1/models.

404 unsupported_endpoint. The base URL is missing /v1, or has an extra path. Use https://tokens.bd/v1 exactly.

An error about tool calling or streaming. The model does not support one of them. Pick a model that does (Streaming, Tool calling).

403 monthly_spend_cap_exceeded or 402 insufficient_credits. The key's cap or your balance is used up. Raise the cap or top up in billing.

429 rate_limited or concurrency_limit. The agent sent requests faster than your limits allow. Wait for Retry-After; see Rate limits.

For the full list of codes, see Errors. Other problems: Troubleshooting.