Docs / Start
Providers and API keys
A provider is the company whose AI model answers you. Shard supports three kinds:
- Your own API key (you pay the provider directly): Z.AI, OpenRouter, Amazon Bedrock, OpenAI, Anthropic, Gemini, Kimi (Moonshot), Meta Model API, Grok (xAI) and Groq.
- A connected account (you use a subscription you already have): Codex (ChatGPT) and Claude Code. See Connected accounts.
- Shard-hosted models (you pay Shard in credits): the Shard provider. See Shard-hosted models and credits.
Pick a provider in Settings → AI → Providers. The list and the models in it come from Shard’s online catalog, so they can change without a plugin update.
Adding an API key#
- Open Settings → AI and pick the provider.
- Paste your key into the API Key box.
- Click outside the box. It shows “Key saved!”.
Next time you open that provider, the box shows “Key set!”. To replace the key, paste a new one the same way. Each provider keeps its own key.
Your key is saved in Studio’s plugin settings on your computer. It is also sent to Shard’s server with every message, because the server makes the call to the provider for you. Shard doesn’t show the key again after you save it. See Privacy and data.
Where to get each key#
| Provider in Shard | Where to get a key | Notes |
|---|---|---|
| OpenAI | platform.openai.com/api-keys | Uses the OpenAI Responses API. Some models offer a Reasoning mode with a “Pro” option. |
| Anthropic | platform.claude.com, then Settings → API keys | The only provider with the Prompt caching switch (see below). |
| Gemini | aistudio.google.com/apikey | Google AI Studio key. |
| OpenRouter | openrouter.ai/keys | One key for models from many companies. |
| Groq | console.groq.com/keys | |
| Grok (xAI) | console.x.ai | |
| Kimi (Moonshot) | platform.moonshot.ai | Uses the international api.moonshot.ai endpoint. |
| Meta Model API | dev.meta.ai, then the API keys tab | |
| Z.AI | z.ai, then the API Keys page | Shard calls Z.AI’s GLM Coding Plan endpoint. TODO (owner) confirm whether a plain pay-as-you-go key works, or a Coding Plan subscription is required. |
| Amazon Bedrock | AWS Console → Amazon Bedrock → API keys | See Amazon Bedrock below. |
TODO (owner) check these links before publishing. They point to each provider’s own site and may move.
Reasoning#
Many models can “think” before they answer. When the model you picked supports it, Settings → AI shows a Reasoning picker.
- Default sends nothing extra, so the provider uses its own default.
- The other options depend on the model, for example Off, Minimal, Low, Medium, High, Extra high and Max.
More reasoning usually means better answers on hard tasks, but slower replies and more output tokens (so a higher bill on your own key).
Some OpenAI models also show Reasoning mode, with Default and Pro.
Your reasoning choice is saved per chat, per provider and per model. When no chat is open (New chat), it becomes the default for that model. The chat header shows your choice, for example “High reasoning”.
Prompt caching#
When you pick Anthropic, Settings shows a Prompt caching button. It is on by default.
- ON: Shard marks the stable start of the conversation so Anthropic can cache it. Cached input is cheaper on later requests.
- OFF: Shard stops adding cache markers. Click the button again to turn it back on.
The switch only exists for Anthropic. You can see how much caching helped in the chat’s Chat details → Prompt cache card (see Chatting).
Amazon Bedrock#
Shard uses a Bedrock API key (the kind AWS sends as a bearer token), not an access key ID and secret.
- In the AWS Console, open Amazon Bedrock → API keys and create a key.
- In Shard, pick Amazon Bedrock and paste the key into API Key.
- Make sure your AWS account has access to the Claude models you want to use.
Things to know:
- Shard calls Bedrock in the us-east-1 region. There is no region setting in the plugin yet. TODO (owner) the backend accepts a region, but the plugin UI never sets one.
- The Bedrock models use AWS’s global cross-region inference profiles.
- Claude Fable 5.1 on Bedrock needs AWS’s
aws_reviewdata-retention mode. In that mode, AWS may keep prompts and completions for up to 30 days for review.
Hidden providers#
Some providers exist in the code but are hidden in this version, so they don’t appear in the list. If you picked one in an older version, Shard keeps your saved key and choice, but you need to pick a visible provider to keep chatting.