Model Providers
Add each AI provider once. Chat, agents, and every app behind the LLM Proxy can use it, and none of them sees the key. Paste an OpenAI key, sign in with your ChatGPT subscription, or point at your own vLLM server.

Connect a Provider
Connect with an API key that your company pays for by use, or with a subscription that a person already pays for.
With an API Key
Paste a key, and Archestra tests it and loads the provider's models.
- Go to Model Providers in the sidebar and click Add API Key.
- Pick the Provider and paste the API Key.
- Optionally open Advanced to mark the key Primary, set a Base URL for a proxy or self-hosted endpoint, or add Extra HTTP headers that every request to the provider carries.
- Click Test & Create. Archestra calls the provider with the key and syncs its models.
The new key appears in the Provider API keys table as Configured, and its models appear under Models. To choose who can use it, edit the key and add people or teams under its permissions.
You can also set a key through an environment variable, ARCHESTRA_CHAT_<PROVIDER>_API_KEY (for example ARCHESTRA_). It is the fallback when no stored key applies.
Keys that come from deployment settings, such as Vertex AI, show Env Vars under Storage. Their credentials change only in the environment, and they cannot be deleted. Edit one to choose who can use it.
To hide a provider your company does not use, go to Settings → LLM → Model providers and switch Available off. It leaves every picker, and nobody can add a key for it. Its existing keys keep working. On the same page, you can give a provider a display name. For GitHub Copilot, you can also turn off only personal subscriptions or enterprise plans.
With a Subscription
Already pay for ChatGPT, GitHub Copilot, Microsoft 365 Copilot, or SuperGrok? Use it instead of a metered API key. On Model Providers, click Connect on the provider's card and sign in.
These credentials belong to one person and cannot be shared. An agent that uses one always runs on the chatting user's own subscription. A user who has not connected sees a sign-in prompt in chat. If the vendor rejects the sign-in later, click Connect again; the credential keeps its model and agent selections.
- ChatGPT: first turn on Enable device code authorization for Codex in ChatGPT under Settings → Security. It is off by default, and ChatGPT blocks the sign-in until you turn it on. If the backend's outbound traffic is restricted, also allow its hosts.
- GitHub Copilot and Microsoft 365 Copilot have their own pages: GitHub Copilot, Microsoft 365 Copilot.
ChatGPT and SuperGrok use bills $0 in Costs. GitHub Copilot and Microsoft 365 Copilot use counts as metered. See Subscription vs Metered Cost.
Which Key Pays for a Chat
When several keys exist for one provider, the most specific one wins. Your team shares an OpenAI key, and you add your own. Your chats now run on yours, so your use stops counting against the team's key.
Chat and agents take the first match:
| Order | Key | Example |
|---|---|---|
| 1 | The key picked in the conversation | You switch a chat to the research team's key. |
| 2 | The key set on the agent | A support agent always bills the support team's key. |
| 3 | Your own key | You added a personal Anthropic key. |
| 4 | A key shared with your teams | Platform Engineering shares one OpenAI key. |
| 5 | A key shared with the organization | The company default key. |
| 6 | The ARCHESTRA_CHAT_<PROVIDER>_API_KEY environment variable | The key set at install time. |
When one level has several keys, the Primary key wins, then the oldest one.
- A subscription is used only by the person who connected it, even when an agent or a conversation picks it. Anyone else uses their own subscription, or gets a sign-in prompt.
- For vLLM, Ollama, and Azure, each key points at its own server. A key whose server does not host the model you picked is skipped for one that does.
- Clients that call the proxy pick their key through their authentication method.
Supported Providers
Each provider has its own proxy URL, https://<archestra-host>/v1/<path>, with the provider's own API. A client that works with the provider works with the proxy. Providers marked Router also work through the Unified Model API. A linked name has a setup page.
| Provider | Path | APIs | Router | Notes |
|---|---|---|---|---|
| Amazon Bedrock | bedrock | Converse, InvokeModel | Yes | API key, AWS access keys, or IAM role |
| Anthropic | anthropic | Messages | Yes | Also Claude on Microsoft Foundry and Vertex AI |
| Archestra | archestra | Chat Completions | No | Another Archestra instance as the upstream |
| Azure AI Foundry | azure | Chat Completions, Responses, Embeddings | Yes | API key or Microsoft Entra ID |
| Cerebras | cerebras | Chat Completions | Yes | |
| Cohere | cohere | Chat | Yes | |
| DeepSeek | deepseek | Chat Completions | Yes | |
| GitHub Copilot | github-copilot | Chat Completions, Responses | Yes | Personal sign-in or organization GitHub App |
| Google Gemini | gemini | Generate Content, Embeddings | Yes | Google AI Studio or Vertex AI |
| Groq | groq | Chat Completions | Yes | |
| Jev | jev | Decisions | No | Scores content, such as a tool call. It does not chat. See Jev. |
| Kimi (Moonshot AI) | kimi | Chat Completions | Yes, Chat Completions only | Kimi API Platform keys. China endpoint: set ARCHESTRA_ to https://api.moonshot.cn/v1 |
| Kimi Code | kimi-code | Chat Completions | Yes, Chat Completions only | Kimi Code membership keys (sk-kimi-...), which use the plan's quota. China endpoint: set the key's Base URL to https://api.kimi.com/coding/v1 |
| Microsoft 365 Copilot | microsoft-365-copilot | Chat Completions | No | Personal sign-in only, no tools |
| MiniMax | minimax | Chat Completions | Yes | Text only |
| Mistral AI | mistral | Chat Completions, Embeddings | Yes | |
| Ollama | ollama, ollama-native | Chat Completions, Embeddings, native Chat | ollama only | No key needed |
| OpenAI | openai | Chat Completions, Responses, Embeddings | Yes | ChatGPT subscription through Connect |
| OpenAI-compatible servers | vllm | Chat Completions, Embeddings | Yes | vLLM, llama.cpp, LM Studio, SGLang, and others |
| OpenRouter | openrouter | Chat Completions, Embeddings | Yes | See OpenRouter Free Models |
| Perplexity | perplexity | Chat Completions, Responses | Yes | sonar models take no tools; vendor-prefixed models such as anthropic/claude-opus-5 do |
| xAI (Grok) | xai | Chat Completions, Responses | Yes | SuperGrok subscription through Connect |
| Zhipu AI | zhipuai | Chat Completions, Embeddings | Yes |
Each provider's default endpoint can be changed with its ARCHESTRA_<PROVIDER>_BASE_URL variable, listed under LLM Provider Configuration.
OpenRouter Free Models
OpenRouter's :free model variants cost nothing, though they still need an OpenRouter API key. The providers behind them may use your requests to train models, so do not send sensitive data. The openrouter/free model picks a free model for each request. When you add an OpenRouter key and the organization has no default model, Archestra makes openrouter/free the default.
Jev
Jev is TypeSafe's decision model. It scores content—such as tool calls, Guardrails checks, or sessions for Auto Routing—and does not chat. Send requests to https://<archestra-host>/v1/jev/decisions with Authorization: Bearer <your-api-key>.
To use Jev through OpenRouter, set the key's base URL to https://openrouter.ai/api/alpha/decisions. To change the default endpoint, set ARCHESTRA_.
ChatGPT Subscription Egress
If you restrict the backend's outbound traffic, allow HTTPS (port 443) to these hosts. Without them, ChatGPT sign-in and chats fail. Allowing api.openai.com does not cover them.
| Host | Used for | Setting |
|---|---|---|
auth.openai.com | Sign-in and token refresh | ARCHESTRA_ |
chatgpt.com | Chat requests and model lists, at /backend-api/codex | ARCHESTRA_ |
- Allow them by hostname, not IP address. Both are served from a CDN, so their addresses are shared and can change.
- If you change either setting, allow the host and port it points to instead.
- People who sign in also need browser access to the sign-in host to enter the code.
Model Pricing, Limits, and Modalities
Archestra fills in each model's prices, context window, and input types for you. It syncs them from the provider and a public model registry. A self-hosted or very new model can have gaps. Set any value yourself, because each one changes how Archestra treats the model:
| Detail | What it changes |
|---|---|
| Pricing | The cost Archestra records for each request. Wrong prices give wrong cost reports and budgets. |
| Context window | When chat compacts a long conversation. |
| Max output tokens | How long one answer can be. Without it, a turn asks for 8,192 tokens, which can cut a long answer short. |
| Modalities | Which files chat sends to the model, and which models you can pick for embedding and OCR. A file the model cannot read goes to the conversation's Files panel instead. |
For an OpenAI connection to a gateway, model refreshes import its published output-token limits. Limits stay with each connection, so gateways can use the same model name with different limits. These connection limits are applied to requests but are not shown on the Models page, which shows the shared model values and manual overrides.
A manual Max output tokens value can lower a connection's budget but cannot exceed its published limit. If no limit is available, set Max output tokens yourself.
To change one, go to Models, edit the model, and open its Pricing, Limits, or Modalities tab. Your values stay through model refreshes. Clear a field to use the synced value again.
