MaxwellAI Cognitive OS.

Before any agent acts, Aurora answers a single question — who is acting, and with what authority?

Aurora is the OS access layer. It models every actor — human employee, AI agent, partner system, citizen — as a first-class identity, with namespaces, scopes and policies. Bring your own identity provider; Aurora speaks OpenID Connect, SAML, SPIRE and passkey natively.

Tenants are isolated as FireFly namespaces, so a single deployment can run many organisations, jurisdictions or trust levels without leaking context. Every authorization decision becomes an audit event the rest of the OS can reason about.

Identity is no longer a login screen. It is the grammar that decides what an agent is authorized to think about.

Magna is the digital constitution — the charter, the treasury and the votes that decide what the organisation is for.

Most organisations run on PDFs, Slack threads and tribal knowledge. Magna encodes the organisation itself as a living charter: articles of constitution, members (human and synthetic), proposals, votes, treasury allocations and runway.

When an agent requests budget, hires another agent or signs a partner contract, the request flows through Magna. The result is an organisation able to prove what it decided, when and by whose authority — without any single person becoming a bottleneck.

If Aurora answers “who”, Magna answers “why”. Together they make the organisation legible to itself.

Atrium is the bridge — the dashboard, the CLI, the voice command and the phone number humans actually touch.

Every Plane must be operable by a person, on demand. Atrium is the surface layer: a 60fps mission-control GUI, a scriptable CLI, a voice assistant and a real phone number that routes into the OS. Slack, Teams and chat channels are first-class peers.

Goals, budgets, tickets, pulses, org charts — anything living in other Planes can be inspected, commanded or escalated through Atrium. It is the only Plane most humans need to learn, because it talks to all the others.

An agentic organisation should be reachable by any means — not just a web app.

Froid is planning without improvisation — specialised agents turning vague intent into executable specification, phase by phase.

Most teams ship features on a hunch. Froid imposes rigour on delivery: an Analyst frames the problem, a Product Manager writes the stories, an Architect defines the contracts, a Developer implements, a QA certifies. Each agent gets only the context it needs, when it needs it — cold, scoped, auditable.

Built on DDD (Domain-Driven Design) and TDD (Test-Driven Design) conventions, Froid turns a vague request into a specification, a sprint board and an executable task list — without losing track of who asked for what.

Vibe-coding warms up the hype. Cold planning scales to a thousand agents.

Flow is the visual canvas where LLMs, tools, databases and agents are wired into executable workflows.

Flow is a low-code plane for agents and flows. Inputs (chat, voice, webhook, SQL), models (Nemotron, Anthropic, Ollama), memory (working, semantic, episodic), tools (MCP, Python, web search), logic (routers, loops, policy gates) and outputs (chat, A2A, write-back) snap together into testable flows.

Every flow is published as an API endpoint, an MCP server or a JSON contract. RAG and the MCP protocol are first-class citizens. The interactive playground replays flows step by step so a human reviewer sees exactly what the agents did.

If Froid is the score, Flow is the orchestra. Every node is auditable, every edge is typed.

Jarbas is the harness where agents are built, operated, governed and improved — at scale, in production.

Once a flow is published, the agent needs a runtime contract: which model, which policies, which channels, which slice of traffic, which approvers. Jarbas is that contract. Rust runtime, integration and trust layers in Go, cognition in Python, UX in TypeScript — every agent has a manifest, a sandbox and a healthcheck.

Jarbas runs the lifecycle: ingest data, distil, evaluate, simulate, execute, promote. NVIDIA NeMo handles the inference cycles; Hyperledger FireFly coordinates multi-party trust. The MaxwellAI Marketplace makes plugins, agents and policies installable as packages.

An agent in production is not a prompt — it is a deployment, a contract and a chain of custody.

Glue is the bridge to the rest of the world — connectors, webhooks and BYOA adapters moving events at scale.

An agentic organisation is only as effective as the signals it can see. Glue is the real-time event layer: connectors for enterprise APIs (Salesforce, Stripe, Twilio, Zendesk, Mailgun), inbound and outbound webhooks, and BYOA adapters for partner agents speaking A2A or MCP.

Every event is routed, logged, dead-lettered and projected into the appropriate Planes. Asynchronous delivery at scale, with HMAC-signed payloads, exponential backoff and traceability from any event back to its origin. And every event is sealed by FireFly into a tamper-evident chain.

The world doesn't pause for your batch jobs. Glue makes the OS event-driven by default.

Trust is the conscience — approvals, provenance and forensic projections for every consequential act.

Agents are powerful precisely because they act. Trust is the layer that decides which actions require human consent, which may proceed autonomously under policy, and which must be reversible. Approval workflows connect to trusted event streams, sealed by FireFly into a tamper-evident blockchain.

Provenance and audit projections record the entire operational history — who proposed, who approved, what data they saw, what decision they made, what the outcome was. Compliance teams query the same chain for SOC 2, ISO 27001 and forensic analysis.

An ungoverned agent is a liability. An agent under Trust is an incorruptible colleague.

Mind is the unified context layer — document, stream, vector, temporal and relational in a single store.

Most AI systems hide their amnesia behind specialised prompting. Mind solves the problem at the substrate. Built on SurrealDB and Spectron, it holds facts, embeddings, episodes, procedures and preferences in one store, queryable through SurrealQL.

Spectron extends the store into five memory layers: Working (current session), Semantic (facts and ontology), Episodic (past interactions), Procedural (learned patterns) and Preference (user feedback). Agents retrieve the context of a year-old conversation in milliseconds.

An agent without memory is a stranger every morning. Mind is why the OS gets better the longer it runs.

Learner is the training plane — NeMo, Nemotron, voice and customisation for the organisation's own models.

Generic models are a starting point, not a strategy. Learner is the studio where the MaxwellAI Cognitive OS fine-tunes, distils and aligns the organisation's proprietary models on NVIDIA NeMo. SFT, LoRA, DPO and RLHF are first-class workflows; NeMo Curator handles the data; NeMo Guardrails handles safety.

Voice, vision and language live here: Parakeet ASR for transcription, Bark/Riva for speech synthesis, Nemotron for chat. Every checkpoint becomes a NIM microservice that Runner can deploy, with model cards guiding the decision to ship or not.

Owning your model is owning your reasoning. Renting it is renting your competitive position.

Runner is the execution layer — Kubernetes, NIM microservices, observability and centralised identity.

Cognition only matters if it executes. Runner is the orchestration layer: multi-region Kubernetes swarms, NVIDIA NIM microservices for accelerated inference, an observability layer for metrics, logs and traces, and a centralised identity & secrets plane on OIDC, SPIRE and Vault.

Every agent, flow and model checkpoint lands here as a deployment with HPA, drift detection, canary rollouts and real-time remediation. When a GPU throttles or a pod OOMs, the operator sees the incident with full context — not a blank dashboard.

The OS is only as reliable as its hardware. Runner is where SRE meets AI.

Avatar is the face of the OS — real-time presence, transcription and a forge for every persona.

Most enterprise software has no face. Avatar is the studio where the MaxwellAI Cognitive OS gains presence: real-time avatars, transcriptions, prompt-driven personas and a Soul Forge for shaping visual identity, voice and style per channel.

When an employee takes calls from Atrium, when a customer meets a retention agent, when a ticket gains a face and a voice — that is Avatar rendering the moment.

Presence is the most human interface the OS can offer. Avatar is how it becomes visible, trustworthy and reachable.