Use your own AI key
Connect Honawa to your own AI provider or gateway: free on our side, you pay the provider. What you need, settings for common providers, and which models work.
When to use your own key
- You already pay for an AI provider such as OpenRouter, Cloudflare, Anthropic or OpenAI.
- You run your own AI gateway or model server.
- You want a model that isn't one of the Honawa plans.
With your own key, requests go straight from the app to your provider. They don't pass through Honawa's servers, and we don't charge anything. The other option is a Honawa plan, where the AI is included.
What your provider needs
Honawa works with any provider that offers:
- An OpenAI-compatible chat API (the
/chat/completionsformat most providers support). - Tool calling. Honawa works by calling your site's tools, so a model without tool calling can only chat.
- Streaming, so answers appear as they're written.
To show images you attach, the model also needs to accept images. Turn on Let the AI see images I attach only for such a model.
Set it up
In Honawa, open Connections → AI model, choose your own provider, and fill in:
- Provider address: the API's base URL (it usually ends in
/v1). - Model: the model's name exactly as the provider writes it.
- API key: your key. It stays in this computer's keychain.
When you click Save, Honawa sends a small test request. If the address, model or key is wrong, you see the provider's error at once.
Settings for common providers
| Provider | Provider address | Model, for example | Key |
|---|---|---|---|
| OpenRouter | https://openrouter.ai/api/v1 |
z-ai/glm-5.3-flash |
An OpenRouter API key |
| Cloudflare Workers AI | https://api.cloudflare.com/client/v4/accounts/<account id>/ai/v1 |
@cf/openai/gpt-oss-120b |
A Cloudflare API token with Workers AI permission |
| Your own gateway or server | Its OpenAI-compatible base URL | Whatever it serves | Whatever it requires |
OpenRouter is the simplest: one key for models from many companies. Set a monthly limit on the key in OpenRouter's settings.
Cloudflare Workers AI: your account id is in the Cloudflare dashboard's sidebar, under Workers AI or on the account home page. Create the token under My Profile → API Tokens with the Workers AI template.
Your own gateway: anything that speaks the OpenAI chat API with tool calls and streaming should work, for example Cloudflare AI Gateway, LiteLLM, or vLLM or Ollama on your own server. We haven't tested each one; if the test request on Save passes and a simple question like "What is my site called?" works, you're set.
Anthropic and OpenAI also offer OpenAI-compatible addresses for their own keys. We tested their models through OpenRouter, not directly.
Which models work
We run every model through the same 22 tasks: WordPress posts and pages, files, and safety checks such as ignoring instructions hidden in a page. A model passes at 85% overall and 100% on safety. Results from November 2026:
Passed, through OpenRouter
| Model | Score |
|---|---|
z-ai/glm-5.3-flash |
100% |
xiaomi/mimo-v2.6-flash |
100% |
minimax/minimax-m3.1-flash |
100% |
qwen/qwen3.8-flash |
100% |
google/gemini-3.8-flash |
100% (slowest, dearest of these) |
anthropic/claude-haiku-5.5 |
98% |
deepseek/deepseek-v4.1-flash |
98% (price varies a lot by provider) |
openai/gpt-6-luna |
93% |
Passed, on Cloudflare Workers AI
| Model | Score |
|---|---|
@cf/openai/gpt-oss-120b |
95% |
@cf/zai-org/glm-4.7-flash |
93% (about 10 seconds a task) |
@cf/moonshotai/kimi-k2.5 |
93% |
@cf/zai-org/glm-5.3-flash |
100%, but see below |
@cf/zai-org/glm-5.3-flash on Workers AI did every task right when we sent one task at a time with pauses. At normal speed, Cloudflare turned most requests away as rate-limited before the first step. If you use it, expect "rate limited" errors on busy tasks; the same model through OpenRouter didn't have this problem.
Didn't pass (don't use these with Honawa): Llama 3.3 70B and Llama 4 Scout (they followed instructions planted in a page), Mistral Small 3.1 (answered without looking at the site), Qwen3 30B, and gpt-oss-20b (often answered with nothing).
A model not on these lists may still work. Try it on a test site first, with a few reads and one small change.
Costs
Your provider bills you. A short task usually sends about 2,500 to 6,500 tokens and gets a few hundred back; on the cheaper models above that's a fraction of a US cent. Long posts and multi-step changes cost more.
Privacy
With your own key, your requests, and what Honawa reads from your site to do them, go to your provider under their terms. Check whether your provider keeps or trains on requests, and choose a setting or provider that doesn't if that matters to you. More: Privacy.