Choose a XiuRouter agent integration

Find the setup path for each client, retain existing providers, and verify real requests and required tools.

This directory matches XiuRouter’s Agent integrations page. Public pages provide placeholder configurations; select an API key and available model in the XiuRouter console, then inspect actual requests. For ChatGPT desktop, choose macOS, Windows, or Linux and configure local Codex outside the app. The official API key sign-in field does not provide an XiuRouter Base URL. See agent development to compare protocol boundaries for Codex, Claude Code, and Cursor.

Three setup methods

First setup: create or select a dedicated key and add the connection as required by the client.
Replace an existing connection: keep the original provider, Base URL, model, and credential source for rollback; add or update only XiuRouter.
Chat-assisted setup: available only when the agent can access the target configuration or terminal. The agent does not read the key or replace restarting the client and creating a new task.

Integration directory

Client	Configuration entry	Existing setup	Chat assistance	Models
ChatGPT	Outside the app; desktop by OS, CLI/IDE by launch environment	Back up, then add a provider	Merge credential-free TOML	Required for first setup; retain an existing matching ID
Claude Code	Claude Desktop, CLI, or VS Code	Record the old connection	Unavailable	Discover from the gateway and select in the client
Cursor	Settings → Models	Record settings before overriding Base URL	Unavailable	Only eligible OpenAI-family models
DeepSeek Harness	Settings → Models	Add or edit `xiurouter`	Unavailable	Fetch available models
OpenCode	`/connect` and `opencode.json`	Update matching credentials; retain other providers	Merge credential-free JSON	Explicit model list
OpenWork	Workspace `.config/opencode/opencode.json`	Merge a workspace provider	Unavailable	Per-workspace models
Cline	Extension settings	Record the old connection	Unavailable	Exact model ID
Continue	`config.yaml` and local secret	Append a model	Merge credential-free YAML	Chat Completions in this guide
OpenClaw	`~/.openclaw/openclaw.json`	Merge provider and model allowlist	Merge credential-free JSON	List models in both provider and allowlist
Hermes Agent	`hermes model` outside the session	Add a Custom endpoint	Unavailable	Select from the endpoint list in Hermes
Aider	Environment and `--model`	Use a separate test terminal	Unavailable	Start with `openai/MODEL_ID`
Open WebUI	Admin Settings → Connections	Add a separate connection	Unavailable	Discover through `/v1/models`
LibreChat	`librechat.yaml`	Append a custom endpoint	Unavailable	Required `models.default`; `fetch` expands the list
Vercel AI SDK	Server code and environment	No separate product-page path	Unavailable	Exact model ID

Model families and reasoning levels

The pricing page groups reasoning levels into one model-family row while the API keeps the exact model IDs that clients may need. Choose a model for an Agent as follows:

For Gemini 3.x, use the base model ID and send reasoning_effort when the client supports it. -low, -medium, and -high IDs remain compatibility inputs; do not maintain them as four unrelated models.
Antigravity Argon low, medium, and high are native upstream tiers. Use gemini-4-argon-low, gemini-4-argon-medium, or gemini-4-argon-high when the tier must be guaranteed; the base Argon name can fall back to high through upstream mapping. Flash can use the base model with a reasoning parameter and read back the actual tier.
Store the exact model ID in the Agent configuration. Grouped pricing display does not change the API key’s model scope or routing.

After changing a model or reasoning level, start a new Agent task and check the request model, protocol, service group, and successful status in XiuRouter’s request verification view. A text reply alone does not prove that tools, editing, or long-context work.

Antigravity Agent setup

Codex, OpenCode, and Cline: use Responses or Chat Completions with gemini-3.8-flash or gemini-4-argon, then pass reasoning_effort when supported.
Claude Code: use Messages with ANTHROPIC_BASE_URL=https://router-api.xiu.ai.
Native Gemini images, editing, and tools: use https://router-api.xiu.ai, x-goog-api-key, and generationConfig; use the exact image model ID gemini-3.1-flash-image.
Clients requiring a fixed tier: use gemini-4-argon-low, gemini-4-argon-medium, or gemini-4-argon-high.

Do not configure the internal /antigravity path. Verify in order: text, streaming, tool round trip, reversible edit, then images. Check model, endpoint, group, tokens, and cost at each step.

Replace an existing connection

Keep the original configuration before adding or updating XiuRouter:

Protocol	Base URL	Authentication
OpenAI Chat Completions	`https://router-api.xiu.ai/v1`	`Authorization: Bearer`
OpenAI Responses	`https://router-api.xiu.ai/v1`	`Authorization: Bearer`
Anthropic Messages	`https://router-api.xiu.ai`	`x-api-key`; some clients use a Bearer token
Gemini GenerateContent	`https://router-api.xiu.ai`	`x-goog-api-key` or the `key` query parameter

Use your XiuRouter key as the authentication value. Keep the old model configuration only if XiuRouter has the same model ID; otherwise copy an exact ID from the current catalog.

Start a new session or task after configuring. A running agent usually retains the environment, provider, and model read at startup.

What chat assistance can do

Codex in ChatGPT, OpenCode, Continue, and OpenClaw can help merge configuration that contains no credentials. Other clients use settings UIs, parent-process environments, server deployment settings, or external wizards.

An agent cannot change the environment of the parent process that started it.
Keep keys out of model context and project repositories.
Hermes /model switches configured providers; run hermes model outside the session to add one.
Cline, DeepSeek Harness, and Open WebUI store credentials in settings, not ordinary project files.
An agent can merge TOML, JSON, or YAML when its sandbox permits access. You still handle credentials, reload the client, and start a new task.

A changed file does not prove the connection switched. Confirm a complete client response and its matching XiuRouter request record.

Clients without dedicated guides, such as Cherry Studio, AionUI, DeepChat, Lobe Chat, or OpenCat, can use the relevant protocol if their current version offers an OpenAI Compatible, Anthropic Compatible, Gemini Compatible, or Custom provider. Verify proprietary tools, file uploads, hosted search, and specific Responses fields separately.

Prepare once per client

Create a dedicated key using the quickstart.
Restrict its models, quota, and expiration.
Copy an exact model ID only if the client needs an explicit model.
Add the connection, retaining any previous setup for rollback.
Start a small new task and inspect the actual request.

Do not experiment with a production application’s key. Existing IP restrictions may reject cloud-originated client requests; the console currently has no editor for that restriction.

Confirm success

A complete integration needs all three:

A full client response, beyond a key-verification check.
A matching successful request showing the key, model, and endpoint in XiuRouter.
The smallest tool operation required by the client working correctly.

A text reply does not establish editing, image, file, hosted-search, or long-context support. Expand testing according to the real task.

Roll back

Keep the original provider or Base URL. If setup fails, restore it instead of repeatedly changing the same key’s scope. Before removing a test provider, check that no project uses it as the default. Revoke or delete unused test keys.

Sources and review date

Reviewed on November 7, 2026. Antigravity native Gemini, image, tools, and reasoning-level paths were checked against production requests. Client settings, files, and request protocols may change after upgrades; use the current interface and official documentation.