AI provider compatibility
Last verified: 19 October 2026, against dashboard-ai.html v1.0 source code, the automated tests in tests/ (results in tests/last-results.json), and a live run with a real OpenAI API key on 19 October 2026.
Key: - Supported: implemented and tested. - Partially supported: works with conditions (see notes). - Not supported: the dashboard doesn't do it, even if the provider can. - Not verified: implemented but not tested end to end with a real account.
| Capability | OpenAI | Claude (Anthropic) | Company/local AI |
|---|---|---|---|
| Bring your own API key | Supported | Supported | Supported (optional) ¹ |
| Configure model identifier | Supported (free text) | Supported (free text, pre-filled) | Supported (free text) |
| Test connection | Supported ² (live key) | Supported | Supported |
| Dashboard analysis (run an agent) | Supported (live key, 19 Oct 2026) | Not verified with a real account ³ | Supported (tested with a local server) |
| Built-in agents | Supported (live key) | Supported in simulated tests ³ | Supported |
| Custom agents | Supported (provider-independent; live run done) | Supported in simulated tests ³ | Supported |
| Import agent pack | Supported (provider-independent) | Supported | Supported |
| Streaming responses | Not supported | Not supported | Not supported |
| Tool / function calling | Not supported | Not supported | Not supported |
| Persistent agent memory | Not supported | Not supported | Not supported |
| Conversation history | Not supported (single message per run) | Not supported | Not supported |
| Set up the view for me (AI replies in a fixed JSON format) | Supported in simulated tests ⁵ | Supported in simulated tests ⁵ | Depends on the model ⁵ |
| Custom endpoint | n/a (fixed: https://api.openai.com) |
n/a (fixed: https://api.anthropic.com) |
Partially supported ⁴ |
| Custom headers | Not supported | Not supported | Not supported |
Notes
1. Sent as Authorization: Bearer <key> only if you enter one. The dashboard asks you to confirm before it first sends a key to a custom address.
2. For OpenAI, the test first lists your account's models, which tells you whether the key and model name are valid, then sends a one-line message.
3. What was tested, and how:
- OpenAI: a real agent run with a live, paid key succeeded on 19 October 2026. This confirms browser access, the request format and answer display. The model used was not recorded.
- Claude: request format built from the documented API, and a real-API error test with an invalid key (401 invalid x-api-key handled correctly). The success path has only been checked against simulated responses. A live run with a real key is still to be done.
4. What a custom endpoint needs:
- the OpenAI Chat Completions format: POST {base}/chat/completions, messages with system and user roles, reply in choices[0].message.content
- HTTPS, except on localhost / 127.0.0.1
- CORS that allows a local file (Origin: null)
- the address added to the safety lock (see AI-SETUP-GUIDE)
- Tested with a local OpenAI-compatible test server: success, missing model, non-JSON reply, and no-CORS cases.
- Not tested: any named product (Ollama, LM Studio, vLLM, Azure OpenAI, company gateways). "OpenAI-compatible" products differ. Azure OpenAI, for example, uses a different address pattern and an api-key header, so it won't work without changes.
- Set up the view for me asks the model to reply with a small JSON object. The dashboard then checks every value against its own list of settings and drops anything that isn't listed. - Tested: with simulated replies, including invented columns, wrong letter case and non-JSON answers. - Not tested: a real model. Smaller local models may not follow the format reliably; the dashboard then says the answer wasn't in the expected format.
What the dashboard sends to each provider
| Claude (Anthropic) | OpenAI / compatible | |
|---|---|---|
| Address | https://api.anthropic.com/v1/messages |
https://api.openai.com/v1/chat/completions, or {your base}/chat/completions |
| Key header | x-api-key, plus anthropic-version: 2023-06-01 and anthropic-dangerous-direct-browser-access: true (required by Anthropic for browser calls) |
Authorization: Bearer … |
| Body | model, max_tokens: 1200, system, one user message |
model, one system and one user message (no length limit set) |
| Cookies / referrer | Not sent (credentials: omit, no-referrer) |
Not sent |
Model names
The kit doesn't fix a model list. Names change, so use the provider's current list:
- Claude: platform.claude.com, models overview. The box is pre-filled with claude-sonnet-5-5, valid when last checked.
- OpenAI: developers.openai.com, models.
Models differ in cost, speed, limits and quality. Any current text model that the provider offers through the endpoint above should work, but only the cases in note 3 have been tested.
Official references
- OpenAI: API docs · Models · Your data (data controls) · API keys · Is API usage included in ChatGPT Plus?
- Anthropic: Get an API key · Models overview · API and data retention · Privacy Center: retention · Why is API usage paid separately?
Also in this guide set: