Security Standard
Built like it has the keys to your business. Because it does.
An assistant that reads your inbox needs more care than a chatbot on a website. Every production install follows the standard below. Controls not yet operational are marked Roadmap.
Architecture
Six layers between the assistant and your systems.
Click a layer to see what it does.
Layer
Dedicated server
Standard: a separate virtual machine for each client. Nothing shared with another client.
Planned hosting is a dedicated DigitalOcean server per client, with a separate control plane for health, jobs and spend.
Controls
The deployment standard, control by control.
01
StandardSeparate environment per client
Each client install gets its own server and container. Client data is never placed next to another business's data.
02
StandardYour credentials, your accounts
Integrations are set up through OAuth apps and API keys created in your own accounts where the provider allows it, so you can revoke any of them without calling me.
03
StandardApprovals before changes
Changes to a system go through the permission level on your plan. Money, deletions and security settings are blocked on every plan.
04
StandardAudit trail
Reads, drafts, approvals and actions are logged with a timestamp and who approved them.
05
StandardSpend caps
Model and tool usage runs under a hard monthly cap you set, and pauses instead of running past it.
06
StandardKill switch
One switch stops the assistant. You can pull it, and so can I.
07
StandardMonitoring
Heartbeat, scheduled job success, expiring tokens, tool failures and spend alerts are checked daily.
08
StandardNo unrestricted execution
The assistant never runs with open-ended permission in a production system. Every tool is allowlisted and scoped.
09
StandardNo consumer AI accounts
Production clients run on business API accounts only, never a personal ChatGPT or Claude subscription.
10
RoadmapNightly backups with restore tests
Nightly backups of memory, jobs and logs, with scheduled restore tests. The automated backup and restore-test tooling for client environments is not built yet.
11
RoadmapOffboarding export and revocation
An export of memory, jobs and logs, credential revocation and environment destruction. Done by hand today. Automated tooling is not built yet.
12
RoadmapRuntime portability
The design avoids lock-in to one model or runtime vendor. Swapping the runtime for a client has not been tested yet.
Permissions
What it may do, by plan.
- Read email, calendar and documentsread
- Reply to a customer emailapproval
- Create or update a CRM record or taskapproval
- Book or move a calendar eventapproval
- Send an invoice reminderapproval
- Run a browser workflow in a web portalapproval
- Issue a refund or move moneynever
- Delete records or emailnever
- Give medical, legal or financial advicenever
- Change passwords, users or security settingsnever
- read
- It can look. It changes nothing.
- draft
- It prepares the reply, update or task. A person sends it.
- approval
- It makes the change after a person taps Approve.
- execute
- It acts on its own, only inside written rules you signed off.
- never
- Blocked on every plan. No setting turns it on.
Execute on Max applies only to actions listed in rules you sign off in writing. Anything not listed stays at Approval.
Implemented and tested today
What is running now: the public demo.
These controls exist and were tested on the demo service only.
- Synthetic data only, with no code path to client or production data
- Per-session, per-visitor and global daily rate limits, tested by pushing past them
- An on/off switch that returns recorded responses when the demo is disabled
- No send, charge or delete tools exist in the demo
Plain limits
What I will not claim.
- No client environment is running yet. The controls above are the NigelBuilds deployment standard, not a record of past operation.
- NigelBuilds does not hold a SOC 2, ISO 27001 or HITRUST certification. I will not claim one.
- HIPAA work is limited. Clinical records are out of scope unless a signed BAA and a written scope cover it. Operations work that does not touch protected health information is the default.
- Model providers process the text the assistant sends them under their business API terms. I pick providers that do not train on API data by default and tell you which ones are in use.