> ## Documentation Index
> Fetch the complete documentation index at: https://continuum-three-olive.vercel.app/docs/llms.txt
> Use this file to discover all available pages before exploring further.

# Grok

> ACP or headless harness sessions via the grok CLI, with a three-rung effort dial and a live weekly credit gauge in Usage.

Continuum integrates Grok (xAI) through the shared harness, over whichever transport the installed CLI supports.

## Requirements

The `grok` CLI must be installed and signed in. Continuum resolves `grok` on your `PATH`; if it is missing, **Settings → Providers → Grok** points you to [x.ai](https://x.ai), and the sign-in action runs `grok login --oauth` in an embedded terminal. Auth is managed by the `grok` binary itself, so "installed" on the Providers screen means the binary is on `PATH`, not proof of a live login.

## How it integrates

Grok 0.2.101 and later speak ACP through `grok agent stdio`. Continuum probes for that capability once per binary (cached on path, size, and mtime) and prefers it when present — ACP carries tool calls and structured messages, not just text.

When the installed CLI doesn't advertise ACP, Continuum falls back to a **headless one-shot transport**, spawning a process per turn:

```bash theme={"dark"}
grok --prompt-file <tmp> --output-format streaming-json --no-leader \
     [--continue] [--model <m>] [--reasoning-effort=<e>] [--always-approve]
```

It parses the resulting NDJSON stream into harness events, so chat, reasoning ("Thinking") rows, turn state, and completion render like the other harness backends. Multi-turn sessions use `--continue`, which resumes the most recent grok session for that working directory; because each session has its own cwd, turns chain correctly. On this path the flag values are passed in `--flag=value` form so a value that looks like a flag can't be parsed as one.

Send, interrupt, and provider selection flow through the same daemon endpoints as Claude and Codex. Grok is a non-Anthropic agent, so Continuum strips Anthropic and Claude credentials from its environment before launch — your Claude token never reaches it.

Grok sessions appear in the [Code tab](/docs/surfaces/code) session list and workbench, and in the [Usage tab](/docs/surfaces/usage) gauge and token history.

## Models

Continuum bundles three Grok models:

| Model                      | Notes                            |
| -------------------------- | -------------------------------- |
| **Grok 4.5**               | Default; coding and agentic work |
| **Grok Build**             | Agentic coding                   |
| **Grok Composer 2.5 Fast** | Fast iteration                   |

These ids go to the CLI's `--model` flag. On the ACP path, the model is passed at launch — Grok takes model and effort only when the agent starts, so changing either mid-session respawns the harness bridge.

## Effort

Grok supports a reasoning-effort control, but a narrower one than Claude or Codex: the dial offers **Low, Medium, and High** only, passed to `--reasoning-effort`.

<Note>
  This is not a Continuum limitation. Verified against grok 0.2.102, `grok-4.5` accepts exactly those three; `minimal`, `xhigh`, and `max` all hard-reject with "unknown effort level … use one of: high, medium, low" and the turn produces nothing at all. A value carried over from another provider is clamped at the spawn boundary — `minimal` folds up to `low`, and `xhigh` / `max` / the Claude-only Ultracode fold down to `high` — so the binary never sees a value it would reject.
</Note>

Because the dial's top rung for Grok is High, a session that has never had an explicit effort picked shows High rather than a misleading Medium.

## Usage and limits

The [Usage tab](/docs/surfaces/usage) shows Grok's live account limit and your token history.

The gauge reads the same weekly credits endpoint grok.com itself uses (`grok.com/grok_api_v2.GrokBuildBilling/GetGrokCreditsConfig`), authenticated with the OIDC bearer the CLI stores in `~/.grok/auth.json`. It returns a credit-usage percentage and a typed current period.

<Warning>
  xAI moved paid accounts to one shared **weekly** pool in June 2026. The older CLI proxy route (`cli-chat-proxy.grok.com/v1/billing`) still answers, but with a *monthly* Extra Usage Credits envelope — reading its reset as the weekly reset is what used to make the Usage card show countdowns like 22 days. Continuum no longer uses it for the gauge.
</Warning>

Historical token analytics come from a separate locally-recorded ledger, kept independent of the live percentage.

## Multiple accounts

Additional Grok accounts get an isolated config root via `GROK_HOME`, plus an explicit `GROK_API_KEY` at spawn. The live gauge reads the auth file of the signed-in CLI, so secondary accounts don't get independent quota readings.

## Broadcast

Grok can participate in the Chat surface's multi-provider [broadcast](/docs/features/broadcast) mode — one prompt to several providers, answers side by side.

<Note>
  The Chat tab is hidden by default across every client while the product is focused on Code, so broadcast is not reachable from the default UI. See [Chat](/docs/surfaces/chat).
</Note>