@memberjunction/ai-inception

MemberJunction AI provider for Inception Labs, whose Mercury models use a diffusion architecture. It ships two drivers:

InceptionLLM (BaseLLM), for the Mercury chat models (Mercury 2, Mercury Edit 2) through Inception’s OpenAI-compatible API at https://api.inceptionlabs.ai/v1. It extends OpenAILLM from @memberjunction/ai-openai, maps the effort level onto Mercury’s instant / low / medium / high reasoning tiers, and does not offer tool calling.
InceptionDecision (BaseDecision), for Mercury Decide on Inception’s Decisions API (below).

Both use the same Inception API key.

Decisions: Mercury Decide

InceptionDecision calls Inception’s Decisions API, POST https://api.inceptionlabs.ai/v1/decisions, which serves mercury-decide, a diffusion-based decision model released on 2026-09-30. In MJ’s catalog it is the Mercury Decide model. The API speaks TypeSafe’s System One format exactly, so the driver extends BaseHostedSystemOneDecision from @memberjunction/ai, which owns the request and answer mapping, the endpoint, the key check and the error messages; the driver adds Inception’s defaults and the classification of the statuses Inception documents.


Endpoint	`https://api.inceptionlabs.ai/v1/decisions`. A URL passed to the constructor wins, then a credential’s `endpoint`. Either is the full request URL.
Key	An Inception API key, sent as `Authorization: Bearer <key>`: the same key `Mercury 2` uses.
Model	`mercury-decide`, the row’s `APIName` and the driver’s default.
Limits	Up to 255 options per Choice and 10 levels per Score (set as the model’s `Decision` limits). No streaming or tools. Inception documents no context window for this API.
Price	$0.02 per million input tokens, a promotional price 50% off the regular $0.04. Output tokens are reported but not charged.

Two routes. The model has two Inference Provider rows:

Vendor	Driver	APIName	Priority
Inception Labs	`InceptionDecision` (this package)	`mercury-decide`	2 (preferred)
OpenRouter	`OpenRouterDecision` (`@memberjunction/ai-openrouter`)	`inception/mercury-decide-20260930` (pinned), 65,536-token context	1

AIDecisionRunner sorts a model’s rows by Priority, highest first, and takes the first one with a credential; on a failover-eligible error it moves to the next.

Configuration. Bind an API Key AI Credential to the model’s Inception Labs Inference Provider row (or to the Inception Labs vendor, which also covers Mercury 2), or set the legacy variable:

AI_VENDOR_API_KEY__INCEPTIONDECISION=...

Usage. Name the model and the vendor through AIDecisionRunner, or call the driver directly:

import { InceptionDecision } from '@memberjunction/ai-inception';

const decide = new InceptionDecision('your-inception-api-key');
const result = await decide.Decide({
    Model: 'mercury-decide',
    State: { message: 'Please refund the duplicate charge on my account.' },
    Questions: {
        billing: { Kind: 'Likelihood', Instructions: 'Is this request about billing?' },
        urgency: { Kind: 'Score', Instructions: 'How urgent is this request?', Levels: ['Can wait', 'Needs attention soon', 'Needs immediate attention'] },
    },
});

As a fallback. Default Decision’s bindings are unchanged, but Mercury Decide, unmeasured in MJ, is among its last power-matched fallbacks once one of its rows has a credential, because the fallbacks follow PowerRank distance from Default Decision’s target rank.

Existing credentials can be enough to reach it: an Inception key bound to the Inception Labs vendor for Mercury 2, or a default API Key credential, covers its Inception row, and an OpenRouter key set for Jev covers its OpenRouter row. On such a host Default Decision can send a decision’s state to Inception once every earlier candidate has failed. To prevent that, set the row’s Status to Inactive.

Errors and failover.

A missing key fails before any request with a NoCredentials error that allows failover.
The message names Inception’s error.message, or a 422’s FastAPI detail[].msg.
A 401 is an Authentication error. ErrorAnalyzer rates it fatal, which stops the failover loop, as for every other driver.
A 402 (billing inactive or quota exceeded) is a NoCredit error and a 422 (an invalid request) a ModelError; the driver sets these itself, so both allow failover. ErrorAnalyzer alone reads a 402 as Unknown and most 422s as a fatal InvalidRequest.
A 404 (model_not_found) is a ModelError and a 413 (request_too_large) an Unknown error; both fail over. So do a 429 and a 5xx.

Class Registration

InceptionLLM, via @RegisterClass(BaseLLM, 'InceptionLLM'). The Mercury 2 and Mercury Edit 2 rows name it.
InceptionDecision, via @RegisterClass(BaseDecision, 'InceptionDecision'). The Mercury Decide model’s Inception Labs row names it.

Dependencies

@memberjunction/ai - Core AI abstractions, including BaseHostedSystemOneDecision
@memberjunction/ai-openai - OpenAILLM, which InceptionLLM extends
@memberjunction/global - Class registration