TSA AI Workforce
Governed Agent Controlplane · by TSA AI Solution Advisory (M) Sdn Bhd
Run AI agents as digital employees — and be able to prove how they behaved.
TSA AI Workforce is a desktop application that lets a company put agentic AI to work on real business tasks while keeping the whole thing inside a management system it can show an auditor. Every agent has its own identity and credentials, every action it takes lands on an immutable trail, and anything that changes an agent's instructions or your compliance record stops at a human for approval.
It runs on your device. The default install uses an on-device PostgreSQL database: no cloud project, no account, no database subprocessor, and no business data leaving the machine unless you deliberately point it somewhere else.
Agentic AI arrived faster than the governance around it. A digital employee that can read your CRM, write to your customer records and act on your behalf is, in governance terms, a member of staff — but most teams deploy one with no identity, no access boundary, no log, and no way to answer the two questions that matter after something goes wrong:
What did it do, and who approved it?
Large enterprises solve this with a compliance department. Everyone else — a 30-person company, a consultancy, a One Person Company (OPC) — gets the same regulatory exposure with none of the apparatus. That gap is the reason this product exists.
TSA AI Workforce packages the apparatus into software you install once:
- Corporates get an AI management system that maps to standards their customers and auditors already recognise, without building it from scratch.
- One Person Companies and small teams get the same controls without hiring a compliance function — the system does the continuous checking, and you make the decisions it puts in front of you.
The goal is not to slow agentic AI down. It is to make it safe enough to actually use on real work.
This is MIT-licensed, and we mean it. Fork it, rebrand it, run it inside your company, build a product on it. No permission needed and nothing to ask us for.
We shipped ISO/IEC 27001, ISO/IEC 42001 and AI Verify because those are the frameworks we needed first. The machinery underneath them is not specific to any of the three. It is a general management system:
- a clause register — any list of requirements, with owners and status
- continuous checks that observe your estate and open findings, then resolve them when the problem stops being observed
- an audit trail nobody can quietly edit
- approval gates where a change stops at a human before it takes effect
- controlled documents with versions, owners and review dates
- internal audits, management reviews and corrective actions on a cycle
Nothing in there knows that the clauses happen to say ISO/IEC 27001. Point it at whatever your business actually has to prove:
| You already run | What you would add |
|---|---|
| ISO 9001 quality, ISO 14001 environmental, ISO 45001 OH&S, ISO 13485, ISO 22000 / HACCP, IATF 16949, ISO/IEC 27701 privacy | a clause catalogue + checks that read your own systems |
| SOC 2, GDPR, PDPA, sector regulation, customer contractual obligations | the same — a framework is just a list of things you must be able to show |
| Your own SOPs, work instructions, forms and records | the document centre, with approval and retention already wired |
| Processes that today live in a spreadsheet and someone's memory | checks that fail loudly instead of a reminder nobody reads |
Where to start in the code:
- Add a standard —
src/lib/compliance/catalogues.ts. Catalogues are provisioned idempotently on every install, in every deployment mode, so a new framework simply exists for anyone who runs your fork. - Add a check —
src/lib/audit/checks.ts. A check is{ id, title, frameworks, run(db) }returning observations with a deterministic key, a severity, the control ids it evidences, and the raw evidence. The engine handles deduplication, sampling a systemic issue rather than burying the register, and auto-resolving what is fixed. - Everything else — documents, approvals, audits and reviews live under
src/app/compliance/.
Two honest notes. This product maintains the register and collects the evidence; it does not certify you — that is an auditor's job, whatever the standard. And the clause text of most ISO standards is copyrighted by ISO, so keep your catalogue to identifiers and your own wording, as we did.
If you build something on this, we would like to hear about it — but you owe us nothing. The whole point is that a 30-person company should not have to build this from scratch.
The product is built around three governance frameworks. It does not certify you — certification is an auditor's job — but it maintains the register, collects the evidence, and continuously checks your estate against each one.
The international standard for managing information security. Published by ISO (International Organization for Standardization) and IEC (International Electrotechnical Commission); the current edition is ISO/IEC 27001:2022, whose Annex A defines 93 controls across four themes — Organizational, People, Physical and Technological.
An ISMS is not a checklist of security products. It is the management system around them: who owns what, what the risks are, what you decided to do about each one, and the evidence that you actually did it.
In this product: all 93 Annex A controls are loaded automatically. The
device scan and the continuous audit engine evidence a subset of them from your
actual machine — disk encryption, firewall posture, listening services, patch
level, credential file permissions, logging — and mark the rest as needing a
human assessment. See /compliance/isms.
The first international management-system standard for artificial intelligence, published by ISO/IEC in December 2023. Where ISO 27001 governs information, ISO/IEC 42001 governs AI systems across their whole life cycle: the policy that authorises them, the roles accountable for them, impact assessments on individuals and on society, data provenance, and monitoring in operation. Its Annex A defines 38 controls.
In this product: every digital employee discovered on your machine is
registered automatically as an AI system under AIMS. All 38 Annex A controls are
loaded, plus the management-system clauses the scanners reference. Changing an
agent's instructions is treated as an AI system change and requires approval.
See /compliance/aims.
AI Verify was created by Singapore. It was developed by the Infocomm Media Development Authority (IMDA) together with the Personal Data Protection Commission (PDPC), and since June 2023 it has been stewarded by the AI Verify Foundation, a non-profit that maintains it as an open-source testing framework and toolkit.
It is not law. It is a way to demonstrate, against a common set of criteria, that an AI system behaves the way you claim it does. This product tracks three of its published criteria sets:
| Criteria set | Origin | What it covers |
|---|---|---|
| Model AI Governance Framework (MGF 2020) | PDPC / IMDA | 11 principles: transparency, explainability, reproducibility, safety, security, robustness, fairness, data governance, accountability, human oversight, inclusive growth |
| Model AI Governance Framework for Generative AI (2024) | AI Verify Foundation + IMDA | 9 dimensions covering the generative-AI supply chain: accountability, data, trusted development, incident reporting, third-party assurance, security, content provenance, safety R&D, public good |
| Agentic AI pillars | This product | 8 pillars for autonomous agents acting on a business: mandate boundaries, HITL gates, agent identity, action logging, tool scoping, memory governance, failure containment, multi-agent delegation |
See /compliance/ai-verify.
Singapore is where this matters most immediately for us — TSA AI Solution Advisory is a Malaysian company operating in the ASEAN market, and AI Verify is the regional reference point for demonstrating responsible AI. The ISO standards apply everywhere.
www.tsa-ai.com/product/tsa-ai-workforce
One installer per platform · the download page detects yours automatically
Start at §1 — Your first agent: what to set up, in order, before a digital employee can do anything.
The same handbook is in the app under Settings → User Manual.
Windows
Run the .exe. If the build you downloaded is not code-signed, SmartScreen may
warn on first launch — click More info → Run anyway.
macOS
Open the .dmg and drag TSA AI Workforce into Applications. If the build is
not notarised, Gatekeeper may require right-click → Open the first time.
Linux
Use the .AppImage — it runs on any distribution without a package manager.
chmod +x ./tsa-ai-workforce-*.AppImage
./tsa-ai-workforce-*.AppImageOne installer gives you two things that talk to each other on your own machine: a desktop shell where you chat with and manage your agents, and a Governed Agent Controlplane — a web console, embedded in the same window, that governs them.
┌──────────────────────── TSA AI Workforce (one app) ────────────────────────┐
│ │
│ Desktop shell (Electron) Workforce tab → Controlplane │
│ ├── Chat with a digital employee ├── Agents + document centre │
│ ├── Profiles (one per agent) ├── ISMS / AIMS / AI Verify │
│ ├── Skills, Tools, Models ├── Internal Auditor │
│ └── Gateway (messaging) ├── Approvals (HITL) │
│ │ └── CRM module │
│ │ ▲ │
│ │ reads + writes │ loopback :17300 only │
│ ▼ │ │
│ ~/.hermes ← agent files on your disk ┌────────┴─────────┐ │
│ (SOUL.md, config.yaml, skills, docs) │ MCP gateway │ │
│ │ │ scope-checked │ │
│ └──── folder sync ─────────────▶│ audit-logged │ │
│ └────────┬─────────┘ │
│ ▼ │
│ On-device PostgreSQL (default) │
│ or your own Supabase / Postgres │
└────────────────────────────────────────────────────────────────────────────┘
A digital employee is a folder: a SOUL.md that defines who it is, a
config.yaml, and whatever skills, prompts and documents you give it. The
Workforce discovers that folder rather than owning it — your agent files
stay where they are, in a format you can read and version-control yourself.
A folder sync inventories every agent and every file that shapes it, recording
path, size, SHA-256 and modification time. File contents are not copied into
the database — the hash is enough to detect drift, and your agent material
stays on your disk. Credential files (.env, keys, tokens) are inventoried so
you can audit that they exist, and their contents are never read.
Each discovered agent is registered automatically as an AI system under AIMS.
An agent never touches the database directly. It calls the MCP gateway, which:
- authenticates it by its own API key (each agent has one, never shared),
- checks the scope required for that resource and operation,
- refuses high-risk writes outright — an agent can propose a change to a compliance record or another agent's instructions, but cannot apply one,
- stamps its identity on anything it creates, so an agent cannot make its own proposal look like a human's,
- writes every call to the immutable audit trail.
Available through it: CRM, the compliance registers, the audit evidence, the agent register and file inventory, the shared work queue, and notifications.
Changes to an agent's SOUL, its config, its files, or your compliance status become change requests. They sit in Approvals until a person with the right role decides. On approval — and only then — the change is written to disk or to the record, with the prior content preserved so the decision is reversible and reviewable.
Deleting a file is included: a management system that cannot delete cannot implement ISO 27001 A.8.10 or retire an AI system. The prior content is captured into the request first, so an approved deletion leaves a record rather than a hole.
A built-in digital employee, provisioned with the product, with its own Hermes profile and its own credentials. It runs continuous checks across the estate, opens findings when something drifts, closes them when it stops reproducing, and raises a desktop notification when something needs you. You can open a chat with it and ask.
It can read everything and approve nothing — including its own proposals.
A scheduled device scan records real endpoint posture — OS and patch level, disk encryption, firewall, listening services, backup recency, credential file permissions — and maps each result to the ISMS/AIMS control it evidences. That turns "we have a policy" into "here is the machine, here is the date, here is the reading".
- Desktop — install the Hermes Agent runtime, pick an AI provider.
- Workforce tab — the Controlplane starts on
127.0.0.1:17300and embeds itself. Create the administrator account (stored on this device). - The database builds itself, the ISO and AI Verify catalogues load, and the Internal Auditor is provisioned.
- Connect your agent folder, or let the app use the Hermes home it manages.
- Mint Agent Keys for the employees that need business-module access.
Each of those steps is a separate gate — installing the app does not give an agent a model, and connecting the folder does not give it permission to touch business data. Download the User Manual (PDF) for the full walkthrough; §1 covers exactly what a brand-new install still needs and in what order.
- One installer — Electron shell + bundled Controlplane + auto-configured MCP
- Zero-configuration database — on-device PostgreSQL; no cloud account required
- Local-first — Controlplane binds
127.0.0.1only; agent files never leave the host - Per-agent identity — separate API keys, scoped by role, individually revocable
- Human-in-the-loop — agents propose, people dispose, every decision logged
- Continuous assurance — ISMS / AIMS / AI Verify checks on a schedule, not once a year
- Agent document centre — browse and edit every file that shapes an agent, through approval
- Optional business module — CRM, reachable by agents through the gateway
- Open-source shell — no proprietary digital employees or SOUL packs are shipped
| Screen | Description |
|---|---|
| Chat | Streaming conversation with an agent, with tools and slash commands |
| Profiles | One Hermes profile per digital employee |
| Workforce | The Governed Agent Controlplane — everything below |
| Agents | Register, SOUL / config editor, document centre, activity |
| Compliance | ISMS, AIMS, AI Verify, documents, management reviews |
| Audit | Internal Auditor, findings, evidence, device scan, reports |
| Approvals | The human-in-the-loop queue |
| Security | Credentials, agent behaviour, autopilot findings |
| CRM | Optional business module agents can work |
| Settings | Company, security, compliance, PDPA retention, Agent Keys, updates |
From a cold install to a signed-in Workforce: the app checks its configuration, walks four setup steps, and hands you the console. Nothing here asks for a cloud account — the built-in database and the Hermes home on disk are the defaults.
 Startup — configuration check before the first frame |
 Setup 1 · Database — built-in Postgres, or point at your own |
 Setup 3 · Administrator — the account that approves changes and mints agent keys |
 Setup 4 · Digital employees — auto-detected Hermes home, scanned into a local inventory |
 Sign in — with the administrator created during setup |
 Dashboard — real-time overview across CRM, agents and compliance |
One queue shared between your team and the digital employees, a live monitor for every agent, and the Agent Hub where conversations, activity, commands and skills are mirrored from Hermes.
 Work Queue — assign either way; anything an agent cannot finish comes back with a reason |
 AI Agents — live monitor, current task, SOUL / config / activity per agent |
 Agent Hub — conversations, agent activity, commands and skills |
 CRM — contacts, leads, opportunities, activities, pipeline |
ISO/IEC 27001, ISO/IEC 42001 and Singapore's AI Verify criteria, each with its own control register. Status changes are proposals until a human approves them.
 ISMS — 93 Annex A controls, evidence export, per-control change proposals |
 AIMS — AI systems, controls, risks and impact assessments |
 AI Verify — MGF 2020, GenAI 2024 and Agentic AI 2026 criteria |
 HITL Approvals — SOUL, skill and compliance-status changes, on the audit trail |
 Document Control — policies, procedures, records, forms, templates |
 Audits — planning, findings and corrective actions |
 Management Review — scheduling, minutes and action tracking |
The Internal Auditor runs continuously: it reads the device and the control registers, opens findings itself, and drafts remediation that never auto-applies.
 Audit Dashboard — coverage by framework, open findings, last automated scan |
 Internal Auditor — device review, L1 / L2 scans, remediation suggestions |
 Findings — nonconformities graded to ISO 19011, raised by continuous monitoring |
 Audit trail — append-only record of commands, approvals, key minting, exports |
 Compliance — one status view across all three frameworks |
 Device scan — L1 passive automatically; L2 active checks need your consent |
 Schedule — plan and manage audit engagements |
 Agents — activities, conversations and heartbeats, filtered by agent and date |
 Risks — register scored likelihood × impact, fed from audit findings |
 Reports — generate and download audit reports as PDF |
 Documents — upload, manage and link ISO compliance documents |
 Security — credentials, agent behaviour, autopilot findings |
 Infrastructure — local runtime, desktop bridge, monitor sync runs |
 Settings — company, security, compliance, PDPA retention, Agent Keys, updates |
| Pillar | Standard | Where |
|---|---|---|
| ISMS | ISO/IEC 27001:2022 — 93 Annex A controls | /compliance/isms |
| AIMS | ISO/IEC 42001:2023 — 38 Annex A controls | /compliance/aims |
| AI Verify | IMDA / PDPC / AI Verify Foundation criteria | /compliance/ai-verify |
| Internal audit | Continuous auditor + device evidence | /audit/* |
Notifications are in-app and desktop-native — no SMTP, nothing emailed out.
git clone https://github.com/Steveser1989/TSA-AI-Workforce.git
cd TSA-AI-Workforce
cp .env.example .env
npm install
cd desktop && npm install && npm run devConsole only, without the desktop shell:
npm run devThen open http://127.0.0.1:17300.
See docs/DESKTOP-PACKAGING.md, docs/USER-MANUAL.md, docs/ARCHITECTURE-DESKTOP.md, supabase/README.md.
Agent data lives under %LOCALAPPDATA%\hermes (Windows) or ~/.hermes
(macOS / Linux).
- Linked — the Hermes home is known to the app
- Agent running — the Hermes Agent process / API is healthy
- Console healthy —
/api/health?live=1responds
MyOPC — www.myopc.io
Malaysia introduced the One Person Company (OPC) structure, letting a single individual incorporate and run a company alone. MyOPC is our platform for that founder: incorporation and the operational back office of running a one-person company, in one place.
MyOPC and TSA AI Workforce address the same person from two directions. MyOPC handles the company; TSA AI Workforce handles the digital employees that company hires — and makes sure a business of one can still meet the standards a business of a thousand is held to.
| Website | www.tsa-ai.com |
| Product page | www.tsa-ai.com/product/tsa-ai-workforce |
| support@tsa-ai.com | |
| Phone / WhatsApp | +60 16 903 7339 |
| Company | TSA AI Solution Advisory (M) Sdn Bhd |
- Upstream shell: fathah/hermes-desktop (MIT)
- Agent runtime: NousResearch/hermes-agent
- Notices: docs/THIRD_PARTY_NOTICES.md
MIT — see LICENSE. Branding and product identity: TSA AI Solution Advisory (M) Sdn Bhd.