Documentation Index
Fetch the complete documentation index at: https://mintlify.com/alibaba/page-agent/llms.txt
Use this file to discover all available pages before exploring further.
Page Agent supports any model that implements the OpenAI API specification and provides tool call support — including public cloud providers, local runtimes, and self-hosted proxies. The list below covers models that have been explicitly tested; any OpenAI-compatible endpoint that supports tool_choice should work.
Tested Models
Models marked with ⭐ are the recommended baseline models: fast and lightweight with strong tool-call capabilities.
Qwen
OpenAI
Anthropic
Google
DeepSeek
Others
| Model | Baseline |
|---|
qwen3.7-max | |
qwen3.7-plus | |
qwen3.6-max | |
qwen3.6-plus | |
qwen3.6-flash | |
qwen3.5-plus | ⭐ |
qwen3.5-flash | ⭐ |
qwen3-max | |
| Model | Baseline |
|---|
gpt-5.5 | |
gpt-5.4 | |
gpt-5.4-mini | ⭐ |
gpt-5.4-nano | ⭐ |
gpt-5.2 | |
gpt-5.1 | |
gpt-5 | |
gpt-5-mini | |
gpt-4.1 | |
gpt-4.1-mini | |
| Model | Baseline |
|---|
claude-sonnet-5 | |
claude-fable-5 | |
claude-opus-4-8 | |
claude-opus-4-7 | |
claude-opus-4-6 | |
claude-opus-4-5 | |
claude-sonnet-4-5 | |
claude-haiku-4-5 | ⭐ |
| Model | Baseline |
|---|
gemini-3.5-flash | ⭐ |
gemini-3.1-pro | |
gemini-3.1-flash-lite | |
gemini-2.5-pro | |
gemini-2.5-flash | |
| Model | Baseline |
|---|
deepseek-v4-pro | |
deepseek-v4-flash | ⭐ |
deepseek-3.2 | |
| Provider | Model | Baseline |
|---|
| MiniMax | MiniMax-M2.7 | |
| MiniMax | MiniMax-M2.5 | |
| xAI | grok-4.3 | |
| xAI | grok-build-0.1 | |
| MoonshotAI | kimi-k2.7-code | |
| MoonshotAI | kimi-k2.6 | |
| MoonshotAI | kimi-k2.5 | |
| Z.AI | glm-5.2 | |
| Z.AI | glm-5.1 | |
| Z.AI | glm-5 | |
| Z.AI | glm-4.7 | |
Model Selection Tips
Prefer fast, lightweight models with strong tool-call (function-calling) capabilities. The ⭐ baseline models offer the best balance of speed, cost, and reliability for typical web automation tasks.
- ToolCall quality matters most. Models with weak tool-call support may return malformed responses. Page Agent has built-in error recovery for common format errors, but persistent failures will cause tasks to fail.
- Avoid very small models. Models smaller than ~10B parameters typically cannot handle the complexity of Page Agent’s tool definitions reliably.
- Context length is critical. A typical rendered page requires 10–20k tokens of context. Ensure your model and plan support at least 16k tokens.
Basic Configuration
// OpenAI-compatible services (e.g., Alibaba Bailian)
const pageAgent = new PageAgent({
baseURL: 'https://dashscope.aliyuncs.com/compatible-mode/v1',
apiKey: 'your-api-key',
model: 'qwen3.5-plus'
});
Free Testing API
A free testing endpoint is available for evaluating Page Agent without a paid API key.
For technical evaluation and R&D only. Data is processed via servers in Mainland China (proxied via Alibaba Cloud FC to BaiLian Qwen models). Do not input any personally identifiable information (PII) or sensitive data. By using this endpoint you agree to the Terms of Use. Never use this endpoint in production. # qwen3.5-plus / qwen3.5-flash
LLM_BASE_URL="https://page-ag-testing-ohftxirgbn.cn-shanghai.fcapp.run"
LLM_MODEL_NAME="qwen3.5-plus"
Production Authentication
If you are building a personal assistant, you can connect directly to your LLM provider. For web app integrations, set up a backend proxy so API keys are never exposed to the browser:
const agent = new PageAgent({
baseURL: '/api/llm-proxy',
model: 'gpt-5.1',
customFetch: (url, init) =>
fetch(url, { ...init, credentials: 'include' }),
});
Never commit real LLM API keys to frontend code. Use a backend proxy with cookie-based or token-based authentication for production deployments.
Prompt Caching
Some LLM providers support prompt caching, which can significantly reduce latency and cost for long system prompts. Because each provider exposes caching through different mechanisms, use the transformRequestBody hook to inject provider-specific cache hints.
Claude supports global automatic prompt caching. When using a Claude-compatible proxy, add cache_control at the top level of the request body:const pageAgent = new PageAgent({
baseURL: 'https://your-claude-proxy.example/v1',
apiKey: 'your-api-key',
model: 'claude-sonnet-5',
transformRequestBody: (requestBody) => ({
...requestBody,
cache_control: { type: 'ephemeral' },
}),
});
Alibaba Cloud Bailian Qwen supports message-level cache_control. Apply it to the system message for maximum benefit:const pageAgent = new PageAgent({
baseURL: 'https://dashscope.aliyuncs.com/compatible-mode/v1',
apiKey: 'your-api-key',
model: 'qwen3.5-plus',
transformRequestBody: (requestBody) => {
const [systemMessage, ...restMessages] = requestBody.messages
if (systemMessage.role !== 'system' || typeof systemMessage.content !== 'string') {
return requestBody
}
return {
...requestBody,
messages: [
{
...systemMessage,
content: [
{
type: 'text',
text: systemMessage.content,
cache_control: { type: 'ephemeral' },
},
],
},
...restMessages,
],
}
},
});
Local LLMs
Page Agent works with local OpenAI-compatible runtimes like Ollama and LM Studio for offline or LAN deployments.
Requirements
Before connecting a local model, confirm all of the following:
- CORS is enabled — the browser cannot call a local endpoint that blocks cross-origin requests.
- Context length ≥ 8000 — a typical page needs ~15k tokens; the default 4k context in most local runners causes truncation.
- Tool call support — the model and runtime must support
tool_choice.
- Model size ≥ 10B parameters — smaller models generally cannot handle Page Agent’s tool definitions.
Ollama
Tested on Ollama 0.15 with qwen3:14b (RTX 3090 24 GB).
macOS / Linux
Windows (PowerShell)
OLLAMA_CONTEXT_LENGTH=64000 OLLAMA_HOST=0.0.0.0:11434 OLLAMA_ORIGINS="*" ollama serve
$env:OLLAMA_CONTEXT_LENGTH=64000; $env:OLLAMA_HOST="0.0.0.0:11434"; $env:OLLAMA_ORIGINS="*"; ollama serve
const pageAgent = new PageAgent({
baseURL: 'http://localhost:11434/v1',
model: 'qwen3:14b'
// no apiKey needed for Ollama
});
If browser-side requests fail, verify that Ollama was started with OLLAMA_ORIGINS="*" (or the origin of your page). Without it, the browser will receive a CORS error.
LM Studio
const pageAgent = new PageAgent({
baseURL: 'http://127.0.0.1:1234/v1',
model: 'qwen/qwen3.5-27b'
});
LM Studio does not support named tool_choice values. Set disableNamedToolChoice: true in your PageAgent config, otherwise tool call requests will fail.