ALESA NOVA is an AI engineering framework that lets teams use agentic AI coding under disciplined governance: guardrails, backups, audit evidence, rollback and consent-led business messaging. It is built for teams that need governance, not just speed.

ALESA NOVA Basic is free and open source under the Mozilla Public License 2.0; it installs into Claude Code from its GitHub repository. Licensed editions add governance for production-critical and regulated work.

Who is ALESA NOVA for?

Technical leads and regulated teams in the public sector and other high-trust organisations who need evidence, audit trails and rollback when using AI to build and operate software. Basic also serves developers, learners and computer labs.

Does ALESA NOVA support on-premises deployment?

Yes. ALESA NOVA Server is an on-premises platform for Malaysian organisations that need data to stay in-house, with local models used by default.

AI engineering framework

AI coding for teams that cannot vibe-code production.

Human command, agent execution, a backup before every change and proof before every “done”. Start free with ALESA NOVA Basic.

Install Basic Talk to us

Free and open sourceMozilla Public License 2.0. Source on GitHub
250 regression testsRun on macOS and Linux, x86-64 and ARM64
Runs on your machineNo account, no telemetry, no network calls

ALESA NOVA Basic: guardrails that act at the moment of risk.

A free Claude Code plugin. When an AI agent is about to do something careless, it blocks the action, backs up the file, or sends the agent back to prove its work.

/plugin marketplace add mdrosli-design/alesa-nova-basic

/plugin install alesa-nova-basic@alesa-nova

Restart Claude Code, then run /nova-doctor to see the guards block on your machine.

View the source on GitHub

Secret-leak guard

Real API keys and tokens stay out of client code, commits and pushes, including AI-provider keys.

Dangerous commands

Mass deletes, DROP and TRUNCATE, force-push and piping a download into a shell are stopped.

Insecure defaults

RLS switched off, TLS checks disabled, wildcard CORS with credentials and production DEBUG are blocked.

Backup before edits

Files are copied before the agent changes them, outside your project, so backups are never committed.

Proof before “done”

If the agent claims it is finished without a passing check, it is sent back once to run one.

Continuity

A project brief that every new session reads first. Save progress with a checkpoint, resume on any machine.

Also included: Coach mode for learners, an AI-use report for coursework and client work, and a self-testing doctor. English and Bahasa Malaysia. The guards read command text, so a deliberately disguised command can slip past them; Basic does not replace code review.

Three rules, applied to every change.

Read before writing.
Inspect the code, schema, logs and running system before anything is changed.
Back up before changing.
A rollback path exists before the edit, not after something breaks.
Verify after changing.
Done means a functional check and evidence you can show, not “the command succeeded”.

From a free plugin to governed delivery.

For teams where AI-assisted change reaches production, personal data or public services.

An engineer and an AI agent at an on-premises operations console

AI Commander

A named technical owner sets the objective, risk tier and approval level for the work.

Agents with roles

Claude, Codex, DeepSeek and local models work in defined roles, with review across model families.

Evidence ledger

Backups, test results, logs and rollback paths become part of the delivery record.

On-prem server

Agent gateway, policy packs, a monitoring console and local models by default, so workload data can stay in-house.

Business messaging that is consent-led and easy to stop.

For WhatsApp Business Platform projects, automation is a governed support layer, not a broadcast tool.

Explicit opt-in and a clear business identity before any conversation starts.
Approved templates where required, and no spammy or misleading messages.
Opt-out is honoured, and sensitive or disputed cases go to a human.
Regulated data, secrets and personal identifiers stay out of message flows that do not need them.

A support agent and an AI assistant reviewing a consent-based messaging workflow

Assurance

Reproducible results, published with their limits.

We show what was tested, how it was tested, and what is still open. No system is bullet-proof, and we do not claim one is.

How we test

15 of 15
adversarial attacks contained on a live model running behind NOVA, while 0 of 5 legitimate tasks were blocked.Live-model benchmark, June 2026. A small, defined test set.
0
medium or higher vulnerabilities in 4,730 automated requests against the server API.Nuclei scan, June 2026. Three informational header notices. Not a full security assessment.
Ed25519
receipts attached to answers by the NOVA On-Prem client, verifiable offline with the public key alone.
250
regression tests in ALESA NOVA Basic, run on macOS and Linux on every push.

Governance is the product.

Any AI that touches citizen or customer data needs consent, limited disclosure and an access trail from the first day.

PDPA by design

Designed with the Personal Data Protection Act 2010 and its 2024 amendments in mind: consent, limited disclosure and access trails. Not legal advice or a certification.

Aligned with Malaysian standards

Controls are designed to meet Malaysian public-sector and data-protection expectations. Alignment by design, not a certification.

Defence in depth

Guardrails, evidence, rollback and audit work together, so changes can be explained and rolled back.

AI governance in Malaysia

Start with Basic. Add governance when the work becomes critical.

For government, enterprise, education and regulated teams that want the power of AI coding without handing production judgment to AI alone.