Not a prompt wrapper. Not a usage policy. An application layer with engraved guarantees — independent of the LLM you choose, with no additional relational protection to maintain.
Terminal
npm install @elysea/core
# sk_test_ key required — developer portal
# Then in your code:
const elysea = createElyseaClient({
baseUrl: process.env.ELYSEA_BASE_URL!,
auth: {
apiKey: process.env.ELYSEA_API_KEY!, // sk_test_ via developer portal
apiSecret: process.env.ELYSEA_API_SECRET!,
},
appId: process.env.ELYSEA_APP_ID!,
countryCode: 'GB',
region: 'EU',
});
const result = await elysea.pipeline.run({ userInput, coreUserId });
// result.response — ethics included, no extra linesUsing Cursor, an AI coding assistant? Copy this.
Integrate ELYSÉA CERVEAU into this project. Install the SDK: npm install @elysea/core sk_test_ key required — developer portal: https://elysea.app/en/developers Initialise the client: import { createElyseaClient } from '@elysea/core'; const elysea = createElyseaClient({ baseUrl: process.env.ELYSEA_BASE_URL, auth: { apiKey: process.env.ELYSEA_API_KEY, apiSecret: process.env.ELYSEA_API_SECRET }, appId: process.env.ELYSEA_APP_ID, countryCode: 'GB', region: 'EU', }); Resolve identity: const { coreUserId } = await elysea.identity.resolve({ userJwt }); Run the pipeline: const result = await elysea.pipeline.run({ userInput, coreUserId }); Display result.response if result.guardianAction === 'pass'.
ELYSÉA Assistant — English version coming soon
The backbone of the SDK. Every exchange passes through a structured sequence: intent parsing → canon verification → LLM generation → ethical post-processing → delivery. You configure the domain canons. The LLM executes within that frame — it cannot bypass it, because the canons are in the conformant execution path, not in the prompt.
Structured storage of user context, isolated by application and by user. Memory is portable (JSON export on request) and erasable (right to erasure, immediate). It only circulates in the ELYSÉA ecosystem with the explicit consent of the user, recorded in the registry — not assumed, not delegated. It is not your asset. It belongs to the person who produced it.
A cryptographically signed asset you deploy in your interface. Your users can scan or verify that your product meets ELYSÉA CERVEAU standards. The label cannot be displayed without an active connection — and you cannot keep it if you violate the canons. It is a permanent audit, not a badge.
Extension points to configure the domain, linguistic posture, local refusal rules and human escalations. What you can configure: the domain, tone, domain limits, redirections. What you cannot configure: absolute prohibitions, ethical canons, AI identification.
The AI cannot impose its will or constrain a choice. It presents options — it does not steer. The user remains in charge of their decision.
No affective re-engagement, no engagement optimisation, no algorithmic retention. The user decides when to return — the AI does not call them back.
The AI identifies itself as infrastructure. In critical zones (emergency, distress, medical or legal decision), it redirects to the qualified human — without exception.
No biased influence, no hidden persuasion, no non-consensual nudging in any generation. This is not configurable.
A third party can tighten. Never loosen.
You can configure ELYSÉA to be stricter than the baseline — block topics, reduce the types of memory used, require more granular consent. You cannot configure ELYSÉA to be less protective of your users. The Guardian SDK rejects any configuration that would weaken a protection — by architecture, not by policy.
You can configure ELYSÉA to be stricter than the baseline — block additional topics, reduce the types of memory used, require more granular consent, restrict posture to your domain. What you cannot do: disable a prohibition, lower a guarantee, weaken an existing protection.
What you can do
What no one can do
ELYSÉA CERVEAU's cross-app memory works at the scale of the entire ecosystem — not just between apps from the same builder. When a user interacts with an app embedding CERVEAU, a structured context builds under their ELYSEAID. That context can be available in any other app in the ecosystem — regardless of who built it — on one condition: the user has explicitly consented to that sharing.
That consent is not an account setting or a checkbox at sign-up. It is a registry entry: which app is requesting access, to which part of the context, for which declared purpose, at what moment. The user can consult that registry, revoke access, export or erase their context. The architecture enforces this — not the builder's privacy policy.
This is what distinguishes a portable identity from a shared database. The context belongs to the person, not to the builders who host it. CERVEAU enforces that distinction in code, at every cross-app access — regardless of the commercial relationship between builders.
A user arriving in your app from the ecosystem brings a structured context — preferences, posture history, working language. With their consent, that context is available from the very first interaction.
Memory does not reset each session. What the user has learned, decided, and built in any ELYSÉA app remains available — and portable — in yours.
If you manage multiple apps, cross-builder memory unifies user context without siloes. The ELYSEAID is the portable identifier shared across the entire ecosystem.
Your users will never receive a clinical conclusion, regardless of how they phrase their question.
No directives to act or orientation toward a treatment — even if the user requests it.
No biased influence, hidden persuasion or non-consensual nudging in any generation.
No affective re-engagement, no engagement optimisation, no algorithmic retention.
Emergency, distress, medical or legal need → automatic redirection to the qualified human.
No prompt from your users — or from you — can modify the pipeline's ethical canons.
Conversations and user memory are never monetised. Not to you. Not to third parties. Never.
The infrastructure presents itself as AI at every relevant interaction. Not configurable by anyone — the Guardian rejects any attempt to suppress it.
The AI Act is in force. Art. 5 since 2 February 2025. Art. 50 since 2 August 2026. This is no longer a calendar — it is applicable law.
Try before you integrate
Two tools available immediately. Submit your own messages to the pipeline before deciding to integrate.
Live comparator
Same message, raw model vs pipeline — side by side.
Open comparator →Playground
Five adversarial attacks + free mode — the pipeline in real time.
Test the pipeline →The builder brings their own LLM model key, from the trial phase onward. Supported providers: OpenAI, Mistral, Gemini, and Claude (Anthropic). Tokens are billed directly by your provider — ELYSÉA does not pay or aggregate them.
The safety classifier is independent of your provider choice
ELYSÉA's safety classifier always runs on ELYSÉA's own keys — never the builder's. The protection is therefore independent of the chosen model provider: whatever key is connected, the safety layer applies with the same independence.
OpenAI
Connect your OpenAI key. Your GPT-4o, GPT-4o-mini and other tokens are billed directly by OpenAI.
Mistral
Mistral AI key — Mistral Large, Mistral Small and other models. Billed directly by Mistral.
Gemini (Google)
Google AI key — Gemini Pro, Gemini Flash. Billed directly by Google.
Claude (Anthropic)
Your personal Anthropic key — separate from ELYSÉA's own integration. You control your Claude tokens directly.
The ethical label follows the allowlist — not just any model.
Connecting your own model via BYOK does not automatically guarantee the ELYSÉA integrity label. The label follows a validated model allowlist — models whose behaviour is being evaluated through the ongoing measurement campaign against the D0 guarantees — results not yet published. If you use a model outside the allowlist, the pipeline functions but the Powered by ELYSÉA badge is not issued.
TEST
Free
Test the full pipeline as long as you need. No payment card.
FOR YOUR PRODUCT
From €2.50
Automatic volume discount. No fixed subscription.
FOR YOUR TEAMS & ON REQUEST
From €30/year
Internal deployment, Enterprise, public sector, education: specific conditions on request.
Two phases. Sandbox: get your sk_test_ key via the developer portal immediately — no credit card, capped at 1,000 MAU. Full pipeline active, Guardian architecture in place. Production sk_live_: individual review of your use case (24-72h) — this is the ecosystem's integrity guarantee for your users.
Define your domain canons, linguistic posture, domain rules and human escalation points. Our engineers accompany you on the first configuration. Result: a versioned, auditable canons file that documents your product's commitments.
The SDK exposes a unified interface, independent of the underlying LLM. You switch models — any leading language model, regardless of version — without touching your application code. The pipeline and canons remain identical whatever foundation you choose.
Activate the integrity label in your interface. Your users can verify it. You cannot disable it without revoking your SDK access. It is not a presentation option — it is the commitment you made to your users, made structurally irrevocable.
sk_test_ immediately via the portal — no credit card. sk_live_ on individual review (24-72h).