Skip to content
Using Odoo Enterprise or Odoo.sh? Support is coming. Enterprise or Odoo.sh? Join the waitlist →
AgentLaireDemoGet it on the Odoo Apps StoreGet it
AgentLaire
Get it

Security

Agents that can’t do more
than your people can

How AgentLaire stays inside Odoo’s security model, what leaves your server, and how you stay in control.

Approval record Update delivery date on S00231, marked as read untrusted content and waiting for approval
Approval record Update delivery date on S00231, marked as read untrusted content and waiting for approval

Shown with AgentLaire Backend Theme

Data

What leaves your server, and when

Requests go only to the model provider you configure, after your redaction rules mask or drop patterns like card numbers and IBANs. With a local model, model requests stay on your own server.

Run step record in Odoo with the IBAN, email and phone redacted before sending to the model
Run step record in Odoo with the IBAN, email and phone redacted before sending to the model

Identity

The agent is the user

Every tool runs with the requester’s rights, record rules and companies. Approved changes run as the approver, after a second check. Tool policies sit on top of Odoo’s security, never instead of it.

Run summary in Odoo: executed as Maya Chen, agent AgentLaire AI Sales, trigger, context record and untrusted flag

Untrusted input

Emails and notes can’t take the wheel

Fields that hold outside text, like email bodies and lead descriptions, are marked untrusted. Once a run reads one, the run is marked untrusted and can’t change anything on its own.

Untrusted means hands off
A run marked untrusted never changes anything on its own, whatever a policy says.
Labeled as data
Untrusted text reaches the model inside delimiters that mark it as data, never as instructions.
Limits on every run
Every run has a tool-call limit and a timeout, and monthly budgets cap the cost, so nothing runs away.
Checked queries
Only allowed fields and standard operators. No raw SQL.

Approval

A person approves every change

By default every change is a proposal you review. A manager can allow one narrow kind in advance, with listed fields, a daily limit and a full log.

Three purchase proposal cards side by side: pending with Approve and Reject, approved with the created RFQ number, and rejected with a reason
Pending purchase proposal card for one vendor, with Approve and Reject buttons
Approved purchase proposal card showing the created RFQ number
Rejected purchase proposal card showing the reason

Around every layer

Audit, MCP and privacy

Audit and control
A log of every run and approval, monthly budgets, and one switch that stops everything.
MCP for AI clients
Connect Claude, Codex or any other MCP client with expiring keys, IP allow-lists, rate limits and a log of every request. Changes over MCP always wait for approval.
Privacy
Nothing goes to a model provider until a manager allows data sharing on it. API keys are never logged, and runs and audit events are deleted after the retention period you set.

Reviewing AgentLaire for a client or your company?

Send us your security questionnaire, or ask anything not covered here.