Skip to content

Provider packages

Some providers are added as packages from the plugin catalog rather than built into Prism. Open Plugins, find the provider in the catalog and choose Install. Most packages are enabled with their own address and need no settings; Azure needs your resource’s address and deployments first, Cloudflare your account’s address. Then add the provider in AI providers with your API key, as you would a built-in one.

Prism stores the key in its sealed vault and sends it only to the provider’s address. The models you can use are read from your account when you connect. The package also declares each model’s context size and whether it streams, calls tools, returns structured output and reads images. A model the provider adds later can be used at once, but without tools or streaming until a newer package version declares it.

Prism keeps installed packages current: after it reads the catalog (shortly after start, then every hour) it installs the newest version this Prism can run of every package you have enabled, with the settings you saved. A version that would send a connected provider’s key to another address or change how it signs in is not installed by itself; Plugins then offers it as an update for you to decide.

When a package declares several providers that take the same key at the same address, you enter the key once. Prism stores it for each of them, a new key replaces it for all of them, and removing the connection removes all of them. AI providers shows them as one connection with all their models. A provider that a package update adds next to a connected one gets the key on its own.

In an organization, a member connects their own account of an installed package on their page, the same way: Amazon Bedrock and Google Vertex AI with their sign-in forms, the others with an API key.

Prism prices a model from its price catalog when the model is listed there. A package may declare a model’s price when no catalog lists the provider (SubQ); a new package version then changes the price without a Prism update. A model with neither has its requests recorded as unpriced.

The chat models of your AWS account in one Bedrock region, Claude included: the system inference profiles (us.…, eu.…, global.…) and the on-demand base models. Install the package, choose the Region in its settings (us-east-1 by default) and add Amazon Bedrock in AI providers. Choose how to sign in: a Bedrock API key (Amazon Bedrock → API keys in the AWS console), or an IAM access key of a user whose policy allows bedrock:InvokeModel, bedrock:InvokeModelWithResponseStream, bedrock:ListFoundationModels and bedrock:ListInferenceProfiles. Prism reads the region’s model list when you connect, so wrong keys are refused then. Long agent conversations are supported. To move to another region, remove the package and install it again.

The Gemini and Claude models of your Google Cloud project through Vertex AI. Create a service account with the Vertex AI User role and a JSON key for it (IAM & Admin → Service accounts → Keys → Add key → JSON), install the package, choose the Region in its settings (global by default) and add Google Vertex AI in AI providers by pasting the whole JSON key; the project comes from the key. A Claude model answers once it is enabled for the project in Vertex AI’s Model Garden. Not every region serves every Claude model: one the chosen region does not serve answers 404, and global serves the most.

Meta’s Muse Spark and Muse Wonder models through your Meta account, signed in the way the Muse Code app signs in. Install the package; it needs no settings. Then choose + Meta Muse Code under Connect an account on the Account quotas page. Prism shows a code and a link to Meta’s page with the code filled in; open it, confirm, and the account appears within a few seconds. Prism then asks Meta for the account’s Muse Code key and sends every request with that key. Meta returns the same key each time and states no lifetime for it, so Prism keeps it until you sign in again. Only the chat models of the account’s catalog are added. The -contributor models are Meta’s discounted mode, in which Meta may use the prompts and answers to train its models. Meta reports no usage limits for the key, so the account has no quota bars. A Muse Code subscription is not required; whether Meta charges requests to a card on the account is Meta’s billing, which Prism does not see.

The model deployments of your own Azure OpenAI or Microsoft Foundry resource: OpenAI models and the other models Azure sells (DeepSeek, Llama, Grok and more). After installing, open the package on the Plugins page and fill in two settings before enabling it:

  • Resource address: your resource’s v1 address, https://<resource>.openai.azure.com/openai/v1 or https://<resource>.services.ai.azure.com/openai/v1. Prism accepts only addresses on these two Azure domains.
  • Deployments: a JSON list of your deployment names as they appear in the Foundry portal, for example ["gpt-prod", "deepseek-r1"]. A name alone means the deployment streams and can call tools. To say exactly what it supports, write an object instead, such as {"id": "gpt-prod", "context_window": 400000, "supports_tools": true, "supports_streaming": true, "supports_vision": true}.

Then add Azure OpenAI / Foundry in AI providers with the resource’s API key. Azure lists no deployments over its API, so connecting does not check the key; a wrong key shows on the first request. To move to another resource, remove the package and install it again.

The Workers AI models of your Cloudflare account. In the Cloudflare dashboard, AI → Workers AI → Use REST API shows your Account ID and creates an API token. Install the package, open it on the Plugins page and fill in the Account address, https://api.cloudflare.com/client/v4/accounts/<Account ID>/ai/v1, then add Cloudflare Workers AI in AI providers with the token. Prism accepts only addresses on api.cloudflare.com.

GLM models from Z.ai’s own API. The package adds two providers: Z.ai (api.z.ai, a key from the Z.ai platform) and Zhipu AI (China) (open.bigmodel.cn, a Zhipu key). The GLM Coding Plan is not supported: Z.ai allows it only in its listed coding tools.

Z.ai’s own model list gives only model names, so the package declares each model’s price (from models.dev, at its most expensive context tier), context length and output limit. A model Z.ai adds later appears without them until the package is updated.

MiniMax models. Two providers: MiniMax (api.minimax.io) and MiniMax (China) (api.minimax.cn).

Open models hosted by Novita AI (DeepSeek, GLM, Qwen, Kimi and others). Novita’s model list is public, so connecting cannot check the key; a wrong key shows as an error on the first request.

Open models served by Parasail. Parasail publishes no public catalog, so its models carry no declared streaming or tool support yet.

Models SambaNova Cloud serves on its own chips. As with Novita, the model list is public and a wrong key shows on the first request.

Vercel’s gateway resells most providers’ models under one key; model IDs carry the maker, such as anthropic/claude-sonnet-5.5. Its model list is public too, so a wrong key shows on the first request.

SubQ’s decision models subq-snap-1.1 and subq-snap-preview through its Decisions API (https://api.subq.ai/v1). They do not write text: for each question about a text they pick one of the answers you describe, with probabilities. Prism uses them as System One judges, for example in the model choice check, and through POST /v1/systemone with the model ext.tm.prism.subq.subq::subq-snap-1.1. They never receive chat requests, so its card has no Test request button. Install the package and add SubQ in AI providers with your SubQ API key. SubQ lists its models without a key, so connecting and Test connection ask it one short decision (about a hundred input tokens) to check the key; a refused key and an account without credits are reported. The package declares the price, $0.05 per 1M input tokens, with no output tokens. From package 1.2.0 Prism takes the models from SubQ’s own list, so a model SubQ adds appears without a new package version; it has no price until a package version declares one.

Inception’s Mercury models (https://api.inceptionlabs.ai/v1): diffusion language models that write a whole answer at once and refine it, which makes them fast. The package declares the chat models Mercury 3, Mercury 2.5 and Mercury 2, with tool calls and structured output, and their prices, since no price catalog lists Inception: Mercury 3 at $0.10 per 1M input tokens and $0.35 per 1M output tokens (Inception’s launch price), Mercury 2.5 at $0.20 and $0.75, Mercury 2 at $0.25 and $0.75. Inception’s model list is public, so a wrong key shows on the first request.

From package version 1.1.0 it also adds Inception Mercury Decide, a System One judge like SubQ: it answers yes/no, choice and score questions about a text, with probabilities. It takes the Inception key, so it needs no connection of its own: mercury-decide appears in the Inception connection, marked “for decisions”, and connecting asks it one short decision to check the key. Choose mercury-decide as the judge of the model choice check, or call it through POST /v1/systemone. Only input tokens are billed, $0.02 per 1M (Inception’s launch price). Mercury Edit and fill-in-the-middle use other endpoints and are not part of the package.