Confidential AI API: Secure LLM Inference at Scale

Beta Request a Router API key through the Contact Us form before configuration. See the Router overview for the quickstart, Models for key-specific model discovery, and API capabilities for tool calling and structured output. Set your API key before you start a coding tool:

export PREM_ROUTER_API_KEY="your-api-key"

Router is not confidential. Do not send sensitive data. During the beta, Router supports Chat Completions only. A tool that requires the Responses API cannot connect yet.

Choose the correct path

Requirement	Use
Sensitive prompts, files, or regulated data	Confidential API
Broader current chat-model catalog	Router
OpenAI-compatible Chat Completions client	Either path, with different credentials and endpoints
Anthropic Messages or Responses API client	Router is not currently compatible

The Router key is separate from the Confidential API key. Do not substitute PREM_API_KEY for PREM_ROUTER_API_KEY or reuse Confidential Proxy KEK settings with Router.

Add Router as a custom provider

In your tool, choose Add provider, then select Custom, OpenAI-compatible, or OpenAI Chat Completions. Enter these values:

Field	Value
Provider name	`Router`
Provider or API type	OpenAI-compatible Chat Completions
Base URL (`baseURL`, `baseUrl`, `apiBase`, or `openAiBaseUrl`)	`https://router.prem.io/v1`
Full endpoint	`https://router.prem.io/v1/chat/completions` only when the tool explicitly asks for a complete endpoint URL
API key (`apiKey`, `openAiApiKey`, or Bearer token)	Paste your Router API key in a GUI; use `PREM_ROUTER_API_KEY` only when the tool supports environment variables
Model (`model`, `modelId`, `apiModelId`, or `openAiModelId`)	Use an exact model ID returned by `GET /v1/models` for the same API key
Streaming	On or automatic
Responses API	Off
Organization or project	Leave blank

Before you configure a model, call authenticated GET /v1/models with the same API key. See Models for the command. When the field is named Base URL or API base, do not append /chat/completions. The final request URL should contain /v1 exactly once. Enter kimi-k3 without a provider prefix; prem/kimi-k3 is only the OpenCode model selector. Select Chat Completions rather than the Responses API. No custom headers are required when the tool has a dedicated API key field. If it only accepts custom headers, add Authorization: Bearer <your-api-key>. Do not enter the literal text $PREM_ROUTER_API_KEY in a GUI unless that tool documents environment-variable expansion.

Streaming requirements can differ by API key. If a model returned by GET /v1/models includes "streaming_required": true, enable streaming or leave it on automatic. OpenCode and Pi support streamed model responses.

Test the custom provider outside your tool with the same connection settings:

curl --silent --show-error https://router.prem.io/v1/chat/completions \
  --header "Authorization: Bearer $PREM_ROUTER_API_KEY" \
  --header "Content-Type: application/json" \
  --data '{
    "model": "kimi-k3",
    "messages": [
      {"role": "user", "content": "Reply with exactly: router ok"}
    ]
  }'

A successful response includes "model":"kimi-k3" and an assistant message.

OpenCode

Create opencode.json in your project root:

{
  "$schema": "https://opencode.ai/config.json",
  "model": "prem/kimi-k3",
  "provider": {
    "prem": {
      "npm": "@ai-sdk/openai-compatible",
      "name": "Router",
      "options": {
        "baseURL": "https://router.prem.io/v1",
        "apiKey": "{env:PREM_ROUTER_API_KEY}"
      },
      "models": {
        "kimi-k3": {
          "name": "Kimi K3"
        },
        "qwen-3.7-max": {
          "name": "Qwen 3.7 Max"
        },
        "qwen-3.7-plus": {
          "name": "Qwen 3.7 Plus"
        },
        "qwen-3.6-plus": {
          "name": "Qwen 3.6 Plus"
        },
        "qwen-3.5-9b": {
          "name": "Qwen 3.5 9B"
        },
        "deepseek-v4-pro": {
          "name": "DeepSeek V4 Pro"
        }
      }
    }
  }
}

Start OpenCode:

opencode

OpenCode uses prem/kimi-k3 as its default provider and model selector. To switch models, use prem/<model-id> with an exact ID returned by GET /v1/models for the same API key. If the response contains either additional model, you can add the matching entries to the models object:

{
  "deepseek-v4-flash": {
    "name": "DeepSeek V4 Flash"
  },
  "qwen-3.8-max": {
    "name": "Qwen 3.8 Max"
  }
}

Add only the entries returned for the active key.

OpenClaw

Add a separate Router provider to ~/.openclaw/openclaw.json. Keep the confidential provider under a different name if you use both:

{
  "models": {
    "mode": "merge",
    "providers": {
      "prem-router": {
        "baseUrl": "https://router.prem.io/v1",
        "apiKey": "${PREM_ROUTER_API_KEY}",
        "api": "openai-completions",
        "models": [
          {
            "id": "kimi-k3",
            "name": "Kimi K3 through Router",
            "reasoning": true,
            "input": ["text"],
            "contextWindow": 1000000,
            "maxTokens": 8192
          }
        ]
      }
    }
  },
  "agents": {
    "defaults": {
      "model": {
        "primary": "prem-router/kimi-k3"
      }
    }
  }
}

The example uses kimi-k3. Confirm that the model appears in GET /v1/models for the active key. If you select another returned model, update both the model id and agents.defaults.model.primary. Set contextWindow from that model’s returned context_length. Run openclaw config validate before starting an agent. Router does not provide enclave confidentiality for the model request, and OpenClaw’s local tools remain outside the confidential runtime.

Hermes Agent

Add a named Router provider to ~/.hermes/config.yaml:

providers:
  prem-router:
    api: https://router.prem.io/v1
    key_env: PREM_ROUTER_API_KEY
    transport: chat_completions
    default_model: kimi-k3
    models:
      kimi-k3:
        context_length: 1000000

model:
  default: kimi-k3
  provider: custom:prem-router
  context_length: 1000000

The example uses kimi-k3. Confirm that the model appears in GET /v1/models for the active key. If you select another returned model, update default_model, both model IDs, and both context-length values. Run hermes config check, then use hermes --oneshot "Reply with exactly: router ok" for the first request.

Goose

Set Goose’s built-in OpenAI provider to the Router endpoint:

export GOOSE_PROVIDER="openai"
export GOOSE_MODEL="kimi-k3"
export OPENAI_API_KEY="$PREM_ROUTER_API_KEY"
export OPENAI_HOST="https://router.prem.io"
export OPENAI_BASE_PATH="v1/chat/completions"

The example uses kimi-k3. Confirm that the model appears in GET /v1/models for the active key. Otherwise, set GOOSE_MODEL to an exact returned model ID. Run a bounded request:

GOOSE_MODE=chat goose run \
  --no-session \
  --text "Reply with exactly: router ok"

Pi

Add this provider to ~/.pi/agent/models.json:

{
  "providers": {
    "prem": {
      "baseUrl": "https://router.prem.io/v1",
      "api": "openai-completions",
      "apiKey": "$PREM_ROUTER_API_KEY",
      "compat": {
        "supportsStore": false,
        "supportsDeveloperRole": false,
        "supportsStrictMode": false,
        "maxTokensField": "max_tokens"
      },
      "models": [
        {
          "id": "kimi-k3",
          "name": "Kimi K3"
        },
        {
          "id": "qwen-3.7-max",
          "name": "Qwen 3.7 Max"
        },
        {
          "id": "qwen-3.7-plus",
          "name": "Qwen 3.7 Plus"
        },
        {
          "id": "qwen-3.6-plus",
          "name": "Qwen 3.6 Plus"
        },
        {
          "id": "qwen-3.5-9b",
          "name": "Qwen 3.5 9B"
        },
        {
          "id": "deepseek-v4-pro",
          "name": "DeepSeek V4 Pro"
        }
      ]
    }
  }
}

Start Pi with Router:

pi --provider prem --model kimi-k3

Run a non-interactive test:

pi -p --provider prem --model kimi-k3 \
  "Reply with exactly: router ok"

Replace kimi-k3 only with an exact ID returned by GET /v1/models for the same API key. If the response contains either additional model, you can add the matching entries to the models array:

[
  {
    "id": "deepseek-v4-flash",
    "name": "DeepSeek V4 Flash"
  },
  {
    "id": "qwen-3.8-max",
    "name": "Qwen 3.8 Max"
  }
]

Add only the entries returned for the active key. supportsStrictMode controls Pi function-tool metadata. It does not disable JSON Schema response formatting.

Get Started

Router

Core Concepts

Handling Errors

Guides

API Reference

Billing

Support

Custom Providers and Coding Tools

Choose the correct path

Add Router as a custom provider

OpenCode

OpenClaw

Hermes Agent

Goose

Pi

​Choose the correct path

​Add Router as a custom provider

​OpenCode

​OpenClaw

​Hermes Agent

​Goose

​Pi

Choose the correct path

Add Router as a custom provider

OpenCode

OpenClaw

Hermes Agent

Goose

Pi