Pi Runtime And Credentials

Router Core includes a first-party runtime built on @mupt-ai/pi-ai. Pi translates normalized router requests into provider API calls and converts output back into the framework’s completion or stream contract. Supports OpenAI, Anthropic, Google, Bedrock, OpenRouter, Fireworks-hosted models, and any OpenAI-compatible API in Pi’s catalog.

Inject Credentials

Single Key
Per Provider

When all models use the same credential:

const pi = await createPiRuntime({
  apiKey: process.env.OPENAI_API_KEY!,
});

A resolver receives provider, model, API dialect, and call purpose:

const pi = await createPiRuntime({
  apiKey: async ({ provider, model, api, purpose }) => {
    const key = await loadApiKey({ provider, model, api, purpose });
    if (!key) throw new Error(`No key for ${provider}`);
    return key;
  },
  timeoutMs: 120_000,
  maxRetries: 2,
});

purpose is execution or selector — use different keys per purpose.

Router Core never reads environment variables. The host reads keys and injects them.

Declare Catalog Models

pi.model() validates a model exists in Pi’s catalog and creates a RouterModel declaration:

const models = [
  pi.model("openai/gpt-5.4-mini"),
  pi.model("anthropic/claude-sonnet-4-6", { defaultReasoningEffort: "high" }),
];
// Omit executor → uses the runtime from createRouter()
const router = createRouter({ models, policy, runtime: pi });

What Gets Translated

Pi carries these normalized features into provider calls:

Messages, images, tool calls, tool results
Prior-turn reasoning and hosted web_search calls
Function tools, named/required tool choice, strict/parallel tools
Temperature, top-p, token limits, stop sequences
Provider-independent reasoning effort and token budgets
Text, JSON object, and JSON Schema output formats
Cache keys, streaming, cancellation

Base64 images work out of the box. Remote image URLs are rejected — download and validate them in the host, or use a custom executor.

Retries And Errors

maxRetries controls Pi’s same-model retries. Provider failures, malformed output, invalid streams, and cancellation become RouterFrameworkError. Cross-model fallback belongs to createRouter, not the runtime: enable it with the fallback option and it works with the Pi executor like any other. For fully managed retries and fallback chains, use Dari’s managed platform.

Per-API Option Gating

Some normalized options can’t be expressed on every provider API. Pi rejects those combinations up front as invalid_request (HTTP 400) instead of surfacing a provider 5xx:

stop sequences on openai-responses / azure-openai-responses
top_p on anthropic-messages
Forcing a specific tool with tool_choice on google-generative-ai / google-vertex
response_format: json_object on bedrock-converse-stream
Non-auto image detail on APIs that can’t preserve it (details are auto, low, or high and pass through on the OpenAI APIs)

Two options degrade instead of failing: json_object on anthropic-messages becomes a permissive JSON Schema (Anthropic has no native JSON-object mode), and metadata is never forwarded — only user, which maps to Anthropic’s metadata.user_id and nothing on other APIs.

Mix Pi And Custom Executors

const models = [
  pi.model("openai/gpt-5.4-mini"),
  { id: "acme/private-model", executor: "acme", capabilities: { streaming: true } },
];
const router = createRouter({ models, policy, runtime: pi, executors: { acme } });

See Custom Executors for the output contracts.

Get Started

Build A Router

Advanced

Pi Runtime And Credentials

Inject Credentials

Declare Catalog Models

What Gets Translated

Retries And Errors

Per-API Option Gating

Mix Pi And Custom Executors

​Inject Credentials

​Declare Catalog Models

​What Gets Translated

​Retries And Errors

​Per-API Option Gating

​Mix Pi And Custom Executors

Inject Credentials

Declare Catalog Models

What Gets Translated

Retries And Errors

Per-API Option Gating

Mix Pi And Custom Executors