GitHub Copilot

GitHub Copilot is the provider OAR starts with: it is always available, and it uses the Copilot CLI that is signed in on the relay host. The relay runs it on one of two engines. The SDK engine is the default and gives you steering, a panel for background tasks and multi-select questions. The Extension engine runs the real Copilot CLI in a terminal you can watch.

OpenAI with your own key runs on the same two engines, so the engine sections of this page apply to it as well. Its key and models are on the OpenAI page.

What you need #

• The Copilot CLI, signed in, on the relay host. Install it with npm install -g @github/copilot, run copilot once and sign in. Your GitHub account needs a Copilot plan that includes Copilot CLI.
• A relay restart after you install or upgrade the CLI. The relay looks for the CLI's runtime when it starts and does not see a later install.
• The GitHub CLI (gh), signed in, for oar copilot, Extension-engine sessions on Windows, and the Copilot card in Check Usage.
• tmux (optional, Linux and macOS) if you want to watch a session's terminal with 🖥️ Inspect tmux console.

The relay does not install or update the Copilot CLI. In Settings → Providers → Copilot, the GitHub Copilot CLI row at the top only reports its state.

How OAR finds the CLI #

The SDK engine drives the runtime that comes with the Copilot CLI. At startup the relay looks in the CLI's package cache and takes the newest version there:

On Linux the two folders follow XDG_CACHE_HOME and XDG_DATA_HOME when you set them. Set COPILOT_PKG_DIR to add a folder to search, or COPILOT_SDK_PATH to name the runtime's copilot-sdk directory directly. If nothing is found, the SDK engine is not available (see Choosing the engine).

Using oar copilot #

oar copilot starts the relay in the background if none answers on its port, then opens the Copilot CLI (gh copilot) in your shell. A relay that it started stops again when the session ends. Arguments after -- go to gh copilot.

oar copilot
oar copilot --port 3400 -- --help

Each run also writes or refreshes a small wrapper in ~/.copilot/extensions/web-relay/, a user-global extension that loads OAR's extension into the Copilot CLI. oar copilot --install-extension does only that and exits (it needs no gh); --no-install-extension skips it for one run. If gh is missing, oar copilot says so and stops; install the GitHub CLI or run only the relay with oar start.

The two engines #

Both engines show the same transcript, question cards, plan boards, model list and per-reply usage line. See Turns and the queue for Steer, Queue and the stuck-turn windows, Relay modes and Compacting for those rows, and ask_user for question cards.

On the SDK engine, Stop ends only the reply. Background agents the turn started keep running under their own Stop in the task panel.

Choosing the engine #

The choice is one setting for the whole relay: Settings → Providers → Copilot → Copilot engine, with the options Extension and SDK — default. Press Save Copilot settings after you change it.

• It applies to conversations whose worker starts afterwards. A session that is already running keeps its engine until its worker restarts.
• If nobody has saved a choice, the relay uses SDK whenever it can and Extension when it cannot. A choice you saved always wins. That includes a saved SDK: if a later restart no longer finds the CLI, the relay does not switch back by itself, and the worker of a new conversation cannot start until the CLI is found again or you save Extension.
• The panel says which engine is set: New Copilot conversations run on the headless SDK worker. No tmux inspector for these sessions. or New Copilot conversations run on the Copilot CLI with the web-relay extension.

To tell which engine a running conversation uses, look at its composer and its tmux console. During a turn the send button reads Steer on the SDK engine (it reads Queue there while a question card or approval is open, the turn is compacting or a tool call is running) and always Queue on the Extension engine. The tmux console shows a Copilot screen only on the Extension engine. The conversation itself does not carry an engine label.

When saving SDK is refused #

The relay refuses to save SDK when it could not run it, and shows the reason in place of the engine description. The stored setting does not change, and the select goes back to the engine the relay actually runs.

Settings #

Under Settings → Providers → Copilot:

Other settings that touch Copilot sessions live elsewhere:

Models #

The relay reads the model list from the installed Copilot CLI runtime a few seconds after it starts. Each model shows the reasoning efforts and the context window the runtime reports. The list starts with auto, then groups models by vendor (OpenAI, Anthropic, Google, xAI, Microsoft, Azure OpenAI, Moonshot AI, then others alphabetically), newest version first.

Until the first discovery, a new relay offers gpt-5.4, gpt-5.4-mini (the default), gpt-5.3-codex, claude-sonnet-4.6 and claude-haiku-4.5.

Use 🤗 Select Models → Copilot in a conversation's ⋯ menu to choose which models appear in the composer, then 💾 Save enabled models. Refresh reads the list again. If the list looks old, press Refresh.

Usage #

• Under a reply, a Copilot turn shows what it used: +N for the credits, with the share of your monthly plan in brackets, and month N% left. A stale tag means the numbers could not be refreshed for that turn. Only Copilot turns get this line; other providers, OpenAI included, do not use Copilot credits.
• In the header, a session line adds up the conversation, for example Session: 1.25 AIC used.
• In Check Usage, open 📊 Check Usage in the ⋯ menu and pick the Copilot tab for the live quota (AI credits or premium requests, chat and plan). On the SDK engine it adds Last SDK worker turn: the credits that turn actually spent, its tokens and model calls. These are one turn's numbers, not a running total, and the section disappears after seven days.

The Copilot card reads your account through gh. If you do not use gh, set GH_TOKEN or GITHUB_TOKEN in the relay's environment.

Importing sessions you already have #

When the relay starts, it imports the Copilot sessions stored on the host, so they appear in your list with their history. It reads them through the installed Copilot runtime. If the runtime is not available, the import is reported as failed. The relay does not guess from the files.

Deleting such a session only hides it: the Copilot CLI's own session stays, because the relay never ran it.

Environment variables #

Set these in the environment of the relay process, then restart the relay.

Limits #

• The relay needs a restart to notice a Copilot CLI that was installed or upgraded after it started.
• On the SDK engine there is no terminal to attach to.
• Messages during a turn queue on the Extension engine; they cannot be steered.
• On the SDK engine the relay gives a turn 2 minutes of silence with nothing running, 5 minutes during a model request and 30 minutes while a tool runs before it fails the turn. A turn that waits on a question card is never treated as stuck. The windows are in Turns and the queue.
• An OpenAI conversation on your own key uses these engines too, but spends your key, so its turns carry no Copilot usage line.

If it does not work #

Agent sessions, relay modes and other behaviour that does not depend on the engine work as described on the pages for sessions and agents together.