Models and providers
To synchronize Mengxi API keys/models, see third-party accounts. Manual provider setup can follow the steps below.
Open
Goal: connect a model and verify the capabilities needed by your task. Prepare the provider's API endpoint, key and exact model ID.
Add a configuration
- Add a clearly named provider in the model/provider settings.
- Choose its protocol. The app includes OpenAI-compatible, Anthropic and Gemini paths; compatible APIs may also distinguish Chat Completions from Responses.
- Follow the configuration screen's base-URL rules. Avoid mixing endpoint formats or duplicating path segments such as
/v1. - Enter the key and load the model list or manually enter a supported ID.
- Select this profile and model in a conversation. Test text first, then a read-only tool call.
Success: text returns and a real tool lists the workspace. Verify a non-sensitive test image separately if you need vision.
Match capabilities to tasks
| Task | Requirement | Check |
|---|---|---|
| Questions | Text generation | Short response |
| Project changes | Tool calling | List and read files before a small edit |
| Screenshots | Vision input or configured supplementary model | Describe details you can verify |
| Device image recognition | Working visual model and device capabilities | Check capabilities and a fresh screenshot |
| Sub-agent work | Enabled orchestration and valid sub-agent configuration | Start with an independent read-only task |
An app capability switch cannot add server-side support. An existing model ID does not guarantee tools, images, reasoning options or long-context support.
Advanced settings
Keep defaults until the connection and basic task work. Adjust temperature, output limit, context or reasoning depth one at a time. Custom headers and JSON must match fields accepted by the server. A small output limit can truncate text or tool arguments; a larger context setting does not ensure server acceptance.
Check pure-prompt mode and Agent tool settings when expecting execution. Web search has its own configuration and backend; a working model connection does not establish search availability.
Diagnose failures
| Symptom | First checks |
|---|---|
| 401/403 | Key, account access and provider restrictions |
| 404 or missing model | Endpoint construction, protocol and exact ID |
| 400 or unsupported field | Custom request fields and reasoning/output options |
| 429 | Provider quota and rate limits; avoid rapid repeats |
| Text works but tools fail | Model tool support, pure-prompt mode, tool switches and arguments |
Keep sanitized errors, protocol and model ID for diagnosis. Never share full keys. Implementation details are in Provider integration.