Skip to main content

Integrating Grok Build with the Turing Platform

Grok Build is xAI's terminal-based AI coding agent. It supports a full-screen TUI, headless command-line runs, and the Agent Client Protocol (ACP), so it can read repositories, run commands, edit files, and manage multi-turn tasks. Because Grok Build accepts custom OpenAI-compatible model endpoints, it can connect directly to the Turing Platform.

Recommended configuration

This guide uses Turing's turing/grok-4.6 through the v1/chat/completions API. Config parsing and model discovery were verified with Grok Build 1.0.5 and the Turing live China-region environment; to verify an actual inference request, use the optional test command below.

Prerequisites​

  • A Turing Platform API key (how to get one)
  • macOS / Linux / WSL2, or Windows PowerShell
  • Permission to set TURING_API_KEY in the shell that starts Grok Build
  • The latest stable Grok Build release is recommended; the fields in this guide were verified against 1.0.5

Step 1: Install Grok Build​

macOS / Linux / WSL2​

curl -fsSL https://x.ai/cli/install.sh | bash

Restart your terminal after installation, or follow the installer's instructions to add ~/.grok/bin to PATH. Verify the installation:

grok --version

Windows PowerShell​

irm https://x.ai/cli/install.ps1 | iex

The PowerShell installer adds %USERPROFILE%/.grok/bin to the user PATH. Open a new PowerShell window and run:

grok --version

Step 2: Prepare a Turing API key​

Grok Build's custom-model configuration uses env_key to read the environment variable containing your API key. Keep only the environment variable name in the config file, not the key itself.

macOS / Linux / WSL2​

Set it for the current shell:

export TURING_API_KEY="your-turing-api-key"

To persist it, add the same line to ~/.zshrc (or ~/.bashrc for bash users), then reload the file:

source ~/.zshrc

You can check whether the variable is set without printing the key:

if [ -n "${TURING_API_KEY:-}" ]; then
  echo "TURING_API_KEY is set"
else
  echo "TURING_API_KEY is missing"
fi

Windows PowerShell​

$env:TURING_API_KEY = "your-turing-api-key"

To persist it as a user environment variable, you can run setx TURING_API_KEY "your-turing-api-key" and then open a new PowerShell window.

Protect your API key

Keep the actual key only in your shell configuration, system environment variables, or a secure secret manager. Do not put it in ~/.grok/config.toml, project files, documentation, screenshots, code repositories, or shared chats; the Grok config file should contain only the environment variable name TURING_API_KEY.

Step 3: Configure the custom model​

Grok Build's user-level config file is ~/.grok/config.toml (or %USERPROFILE%/.grok/config.toml on Windows). Keep any existing [cli], [marketplace], or other settings. If the file already has a [models] table, merge default into that table instead of declaring the TOML table twice. Append the model block below:

[models]
default = "turing-grok-4.6"

[model."turing-grok-4.6"]
model = "turing/grok-4.6"
base_url = "https://live-turing.cn.llm.tcljd.com/api/v1"
name = "Turing Grok 4.6"
api_backend = "chat_completions"
env_key = "TURING_API_KEY"
Quote the dot in the model selector name

[model."turing-grok-4.6"] is the complete TOML table name. Do not write [model.turing-grok-4.6]: TOML splits the dots into nested keys, and Grok Build then treats the final 6 as an unknown field and cannot resolve the default model.

Configuration fields​

FieldPurpose
[models].defaultThe model selector used for new sessions; it must match the model table name below
[model."turing-grok-4.6"]The selector shown by Grok Build; this alias keeps the picker name separate from the request ID
modelThe model ID sent to the Turing API; keep it exactly as turing/grok-4.6
base_urlTuring's OpenAI-compatible root URL; do not append /chat/completions
api_backendRequest protocol; chat_completions maps to /v1/chat/completions and is Grok's default
env_keyThe environment variable containing your API key; this guide uses TURING_API_KEY

Set base_url only through /api/v1. Grok Build appends /chat/completions according to api_backend. Although the Turing backend maps this model to xAI's internal grok-4.6, the client must send the public model ID turing/grok-4.6.

If you want Grok Build to estimate compaction against the 500K context window listed in Turing's model catalog, you may add this optional field to the model table:

context_window = 500000

The field is optional; leaving it out does not affect the verified basic configuration.

Step 4: Inspect the config and start Grok​

First ask Grok Build to show the config sources and discovered models:

grok inspect
grok models

You should see something like:

Default model: turing-grok-4.6
* turing-grok-4.6 (default)

grok models may also report that it could not refresh the official cli-chat-proxy catalog (for example, when no xAI login is present). This does not affect Turing as long as the custom model is shown as using its own API key and is marked default.

To make one minimal end-to-end inference request (this consumes one API quota unit), run:

grok -p "Reply with exactly OK" -m turing-grok-4.6 --max-turns 1 --no-alt-screen

Interactive TUI​

Change into a project directory and start Grok:

cd your-project
grok

You can also choose the model explicitly:

grok -m turing-grok-4.6

Inside an open TUI, switch with /model turing-grok-4.6.

Headless mode​

Use headless mode for scripts, automation, and CI:

grok -p "Explain this repository's architecture" -m turing-grok-4.6

For machine-readable output:

grok -p "Review the current changes for potential problems" -m turing-grok-4.6 --output-format streaming-json

ACP / IDE integration​

Grok Build can run as an ACP agent for an ACP-capable editor or a custom client:

grok agent --always-approve --model turing-grok-4.6 stdio

--always-approve is suitable when an external client owns permission handling. For interactive use, keep Grok's default per-action confirmation.

Model capabilities and request notes​

The Turing model catalog currently lists turing/grok-4.6 as follows:

  • Text and image input, with text output
  • Tool calling and Thinking support
  • v1/chat/completions, v1/messages, and v1/responses
  • 500K-token input and output limits
  • Server-side Web Search is marked unsupported by Turing; do not configure this model as Grok's [models].web_search model

This guide defaults to Chat Completions because it is Grok's default custom-model protocol and the most direct Turing integration. If you switch to responses or messages, verify request-field compatibility with your Grok Build version and the model first.

Prices, cache billing, and model lifecycle can change. Use the live model list as the source of truth. At the time of writing, the catalog shows $2 input, $6 output, and $0.5 cached input per million tokens; requests with ≥200K input tokens use the tiered multiplier, including cached input.

Troubleshooting​

SymptomWhat to do
The default model is missing from grok modelsCheck that the table header is [model."turing-grok-4.6"] and that [models].default matches it exactly; run grok inspect again
grok models says No auth credentials for cli-chat-proxyThis is the official catalog refresh notice; proceed if the Turing model says using its own API key
HTTP 401 / 403Confirm TURING_API_KEY is visible in the shell that starts Grok and contains a valid Turing Platform API key; do not mix a live key with a test URL
HTTP 404Use https://live-turing.cn.llm.tcljd.com/api/v1 as base_url; keep /v1, and do not append /chat/completions manually
Model not found / Model not existThe model field must be the public catalog ID turing/grok-4.6, not the internal upstream ID xai/grok-4.6
Changes do not take effectGrok reads config at session startup; restart Grok, and remember that an existing session may retain the model it was created with

References​

Back​

← Back to AI Coding Tools Overview