Bring your own API key to the SideBoo Chrome extension: OpenAI, Claude, DeepSeek, Gemini, Kimi, GLM, Qwen or OpenRouter. Where to get each key, and how.
8 min read · Last updated 2026-08-27
SideBoo ships no model of its own and proxies nothing. "Ask AI" in the command palette, page summaries, and every other AI feature run on a model service you configure yourself: the key stays in your browser, and requests go straight from your browser to the endpoint you entered.
This page covers three things: how to connect a service, where to get each provider's API key, and what to do when the connection fails.
SideBoo only talks to OpenAI-compatible endpoints. Any service qualifies if it offers these two paths:
| Request path | Purpose |
|---|---|
| {API URL}/chat/completions | Chat. Where generation actually happens. |
| {API URL}/models | Model discovery. Used by "Test service" in settings to check connectivity and pull the model list. |
So the API URL you enter is a base URL and must include the version path (e.g. https://api.openai.com/v1) — not the full /chat/completions URL. Trailing slashes are stripped automatically.
Open the extension's settings page and choose "AI assistant" in the sidebar:
Each provider in the list carries a status badge, so you can see at a glance what's missing:
| Badge | Meaning |
|---|---|
| Needs setup | No API key yet (local services on localhost / 127.0.0.1 are exempt). |
| No endpoint set | The API URL field is empty — only reachable on a custom service. |
| Pick models | Key is in place, but no model is ticked. |
| Ready | Good to go. |
The eight providers below are built-in presets — pick one and the endpoint is filled in, so all you paste is the key.
The API and ChatGPT are separate products — a ChatGPT Plus subscription grants no API credit, you have to add funds on the developer platform. Accounts without a payment method have very low rate limits and hit 429 easily.
Anthropic's preset is labelled just "Claude" in the dialog. Its /models is a native endpoint that expects x-api-key rather than Bearer. SideBoo handles that automatically when you add Claude from the preset. If you instead create a "custom service" and paste the same URL, the connection test will fail on the auth header.
You need to top up in the console before the key has any quota.
Unfunded accounts have a very tight per-minute request limit — consecutive questions will almost certainly error. A small top-up fixes it.
Zhipu AI — the preset carries its Chinese name in the dialog. The path ends in /v4, not /v1 — do not "correct" it out of habit. Lightweight models like glm-4-flash are a cheap fit for summarisation.
Alibaba's Qwen, listed under its Chinese name and reached through Bailian. You must use the compatible-mode entry point; SideBoo does not speak DashScope's native protocol. Enable the model service in the console before first use.
An aggregator — one key reaches hundreds of models, and model IDs carry a "vendor/" prefix. Several free models make it a good way to try providers before committing. The list pulled by "Test service" is long, so use the search box above it.
The trailing /openai is the OpenAI-compatible entry point and cannot be dropped. The service is region-restricted and unavailable from some locations (including Hong Kong). Shared IPs or heavy usage can be flagged as abuse and get the key disabled.
Pick "Custom OpenAI-compatible service" at the bottom of the "Add provider" dialog to connect relay/aggregator gateways, an internal corporate proxy, an Azure OpenAI deployment, or a local inference server.
A custom service needs three fields from you: a name (to tell it apart in the list), the API URL, and the API key.
Take the full chat URL your provider documents and drop the trailing /chat/completions — what's left is the API URL:
https://your-gateway.com/v1/chat/completions ← what the provider gives you
https://your-gateway.com/v1 ← what you enter
Local servers usually don't check credentials, so the API key can stay empty — SideBoo won't block you once it sees the URL points at localhost, 127.0.0.1, or [::1].
Take Ollama: install it, pull a model with ollama pull <model>, and it listens on port 11434 — the OpenAI-compatible URL is http://localhost:11434/v1. The extension holds access permission for http://localhost/* and http://127.0.0.1/* on any port, so plain http:// works.
Ollama also rejects cross-origin requests by default and answers a browser extension with a 403, so set OLLAMA_ORIGINS to chrome-extension://* and restart it:
The variable is read at startup, so the restart is the part people skip. "Ollama, LM Studio and any OpenAI-compatible endpoint" walks through each platform in full.
Click "Add a model ID manually" and type the name from the provider's docs (e.g. gpt-4o-mini). Manually added models are ticked automatically and can be set as the default straight away.
The test reports errors in plain language. Match yours against the table:
| Message | Cause and fix |
|---|---|
| Key rejected (401 / 403) | The key picked up a stray space or newline when copied; the key expired or was revoked; the account lacks access to that model. Generate a fresh key and retry. |
| Endpoint not found (404) | The URL is missing its version path (/v1, /v4, /compatible-mode/v1, …), or you pasted the full /chat/completions URL. For presets, "Restore default" above the API URL field puts it back. |
| Too many requests (429) | You hit the provider's rate limit. Limits on unfunded accounts are usually severe — wait a while or add credit. |
| Can't reach that address | Typo in the URL, service not running, no network — or a local service addressed as something other than localhost / 127.0.0.1 (see the previous section). |
| Connected but 0 models found | The service doesn't implement /models. Add the model ID manually; chat is unaffected. |
| 403 from a local Ollama | Ollama's origin check. Set OLLAMA_ORIGINS as above and restart it — the key is not the problem. |
If the test passes but "Ask AI" still errors, the model ID is usually valid while your account has no access to that particular model — try another one you have ticked.
All charges are billed by the provider you chose and have nothing to do with SideBoo. Pricing models differ (most bill per token, with input and output priced separately), so keep an eye on quota and billing in your provider's console to avoid surprises. Features like page summarisation send an entire article to the model and cost noticeably more per call than a short question — cheap lightweight models (gpt-4o-mini, glm-4-flash, qwen-turbo) are the better value there.
On privacy: