API Providers

Connect to Anthropic, OpenRouter, Ollama, Vercel AI Gateway, or any compatible endpoint

xiantong supports multiple API providers through a built-in preset system. You can connect to Anthropic directly, use aggregators like OpenRouter, run local models via Ollama, or point to any API endpoint compatible with the Anthropic Messages format.

This page covers Anthropic-compatible providers. For Codex/OpenAI connections and multi‑connection setup, see LLM Connections.

Supported Providers#

Setting Up a Provider#

During First Launch#

1. In the setup wizard, select API Key
2. Enter your API key
3. Select a Base URL preset from the dropdown (Anthropic, OpenRouter, Vercel AI Gateway, or Custom)
4. Optionally specify a Model name (required for non-Anthropic providers)
5. The connection is tested automatically before saving

In Settings#

1. Open Settings (gear icon or Cmd+,)
2. Click on the API Connection section
3. Change your API key, base URL, or model as needed

Model Names#

For Anthropic, no model override is needed — xiantong uses its built-in model routing (Sonnet, Opus, Haiku) automatically. For OpenRouter and Vercel AI Gateway, models use the provider/model-name format:

anthropic/claude-sonnet-4
anthropic/claude-opus-4
openai/gpt-4o
google/gemini-2.5-pro
meta-llama/llama-4-maverick

For Ollama, use the local model name directly:

llama3.2
qwen3-coder
deepseek-r1

When the Model field is left empty for non-Anthropic providers, xiantong defaults to Anthropic model name formatting. This works for providers that support Anthropic model names natively but may not work for all providers.

Provider Details#

OpenRouter#

OpenRouter gives you access to hundreds of AI models through a single API key. It handles billing, rate limiting, and fallbacks across providers.

1. Get your API key at openrouter.ai/keys
2. Select the OpenRouter preset in the Base URL dropdown
3. Set your model (e.g. anthropic/claude-sonnet-4)

Browse available models at openrouter.ai/models.

Ollama (Local Models)#

Ollama runs open-source models locally on your machine. No API key is required, and data never leaves your computer. Requirements:

• Ollama 0.14 or newer (for Anthropic-compatible API format)
• A model pulled locally

# Install and pull a model
ollama pull llama3.2

To connect:

1. Select the Custom preset in the Base URL dropdown
2. Enter http://localhost:11434 as the URL
3. Leave the API key empty
4. Set the model name (e.g. llama3.2)

Ollama requires version 0.14+ for compatibility with xiantong. Earlier versions do not support the Anthropic Messages API format. Update with ollama update if needed.

Vercel AI Gateway#

Vercel AI Gateway provides a unified endpoint for routing requests to multiple AI providers with built-in observability and caching.

1. Get your API key from your Vercel dashboard
2. Select the Vercel AI Gateway preset
3. Set your model using provider/model-name format

See supported models in the Vercel documentation.

Custom Endpoint#

For any API that implements the Anthropic Messages format:

1. Select the Custom preset
2. Enter the full base URL of your endpoint
3. Enter your API key (if required)
4. Specify the model name your endpoint expects

This works with self-hosted proxies, enterprise gateways, or any service that implements the /v1/messages endpoint.

Image Input for Custom Endpoints#

Custom endpoints are text-only by default. If your endpoint serves a multimodal model — for example Gemma 4 via Ollama or another OpenAI-compatible proxy — you must opt it into image support explicitly in the connection config. These examples use the low-level LLM connection schema from LLM Connections, where customEndpoint.api selects the endpoint wire format.

Per-model opt-in (recommended)#

{
"customEndpoint": { "api": "openai-completions" },
"models": [
{ "id": "gemma4", "supportsImages": true }
]
}

Whole-endpoint opt-in#

{
"customEndpoint": {
"api": "openai-completions",
"supportsImages": true
}
}

There is currently no automatic capability detection for arbitrary endpoints. If supportsImages is not set, xiantong will treat the endpoint as text-only.

For the full config schema, see LLM Connections.

How It Works#

When you configure a non-default provider, xiantong stores:

• The API key in the encrypted credentials file (~/.xiantong/credentials.enc)
• The base URL and default model in the LLM connection configuration

At session launch, the base URL is passed via the ANTHROPIC_BASE_URL environment variable to the underlying Claude Code SDK.

Troubleshooting#

Connection test fails

Verify:

• The base URL is correct and accessible from your machine
• Your API key is valid and has sufficient permissions
• The endpoint supports the Anthropic Messages API format (/v1/messages)

Model not found errors

Check that the model name matches exactly what your provider expects:

• OpenRouter/Vercel: Use provider/model-name format (e.g. anthropic/claude-sonnet-4)
• Ollama: Use the local model name (e.g. llama3.2)
• Custom: Check your provider’s documentation for valid model identifiers

Authentication errors

• Ensure your API key is correct and hasn’t expired
• For Ollama: no API key should be set (leave it empty)
• Check if your key has available credits/quota

Ollama not connecting

• Verify Ollama is running: ollama list
• Check you’re on version 0.14+: ollama --version
• Ensure the model is pulled: ollama pull llama3.2
• Verify the URL is http://localhost:11434 (note: HTTP, not HTTPS)

Rate limiting

If you hit rate limits, check your provider’s usage limits and consider upgrading your plan or using a different provider.

Security#

Your API key is stored securely in the encrypted credentials file. See Credentials for details on how credentials are protected. The base URL and model name are stored in the LLM connection configuration (not encrypted, as they are not sensitive).