Release v0.113.0
Sep 13, 2026
Remote desktop stops dropping every minute — a server-side bug had been ending healthy sessions since early August, and it is fixed without an agent update. Around that fix: the AI assistant can now write scripts that go through an independent reviewer and land on a readable approval card, service deliverables and key dates for contract work with a customer-facing Documents page, a new Monitors object that replaces hand-assembled alert rules, and a security pass covering Stripe checkout links, integration credentials, alert visibility, and remote-session ordering.
Remote Desktop
- Sessions no longer drop every 30 to 90 seconds. Since early August a background cleanup job had been treating healthy peer-to-peer viewer sessions as abandoned and ending them, because it only recognised one kind of connection as alive. The fix is on the server — no agent update, no action on your side.
- On Windows hosts with two monitors at different scaling levels, the mouse and keyboard now land where you click on the second screen. Previously input and the cursor could be offset or the second display unusable. Requires the updated agent, which arrives on its own.
- Starting and stopping a session are now strictly ordered on both the server and the agent, so a session that was ended can never be revived by a late start command still in flight.
AI Script Authoring
- Ask the assistant to write a script and it produces a proposal rather than running anything. Each proposal is scanned and classified, then reviewed by a separate model that never sees the conversation that produced it, so the reviewer cannot be talked into approving something the author wanted.
- Approval happens on a card that shows the goal, the expected effect, the reviewer's findings and risk tier, the target devices, and the code itself. You can request changes, approve, or reject. After a run, an independent read-back from the device confirms the script did what it claimed before it can be saved to your script library.
- Higher-risk proposals require the scripts permission in the target organization plus MFA to approve, and four-eyes review requests go only to people who hold that permission.
- AI-authored runs show up in the device activity timeline and in a risk dashboard with metrics for reviewed, unattended, and rejected runs, and the script library records how every version was approved.
- An unattended lane exists for organizations that want reviewer-approved scripts to run without a human click. It is off until the MSP raises the ceiling and the individual organization opts in; there is no global switch. Approving several supervised AI actions in a row now takes one passkey or TOTP ceremony instead of one per action.
Service Deliverables & Documents
- Track what each contract owes the customer: deliverables with due dates and key-date reminders, a sweep that flags what is overdue, tickets created from deliverables, and evidence attached automatically when work completes. The AI assistant can query and update them.
- Reusable template sets let you define a standard deliverable list once and apply it to an organization or a contract in one step, including partner-wide sets shared across all your customers.
- Every organization gets a document library with text extraction for search, and documents can be attached to deliverables as evidence.
- Customers see a Service page and a Documents page in the portal, each switchable per organization from portal branding settings.
Monitors & Jobs
- A Monitor is now the one thing a technician writes: a condition, a severity, who gets notified, and what to do about it. Breeze turns it into the alert templates, rules, and automations that already run underneath, so existing alerting keeps working unchanged.
- Monitors have their own page and a tab on configuration policies, and any legacy alert rule converts to a monitor with one click. The AI assistant can create and inspect them.
- Automations now live under a top-level Jobs entry in the sidebar (old links still redirect), Network Monitor is renamed Monitoring, and the device list is now Devices & Assets covering agent devices, manual assets, and network assets in one view. The sidebar shows your last five devices, and Cmd+K opens a jump-back palette of recent devices and pages.
Security Hardening
- A Stripe checkout link could remain payable after you reset the link, recorded a manual payment, voided the invoice, or replaced the Stripe key — a customer paying that stale link would be charged for an invoice that was already settled or cancelled. Every one of those actions now revokes the checkout link and fails if it cannot.
- Discord webhook URLs, Teams client secrets, Grafana, OpsGenie, and PagerDuty keys, and monitoring webhook endpoints were stored in plain text and shown back to anyone in the organization with read access to integrations. They are now encrypted at rest and masked on every screen and response.
- Alert correlation views respected neither site restrictions nor the alerts permission, so a site-scoped technician could see the whole organization's correlated alert graph. Both are now enforced.
- Partner sign-up auto-promotion had silently stopped working — every new partner stayed on probation until someone promoted them by hand. Four separate causes were fixed.
- Organizations that force MFA by role now give a technician who has never enrolled a factor a 14-day enrolment deadline instead of locking them out on their next sign-in. The deadline cannot be extended.
Backup & Recovery
- Bare-metal recovery for Linux: bootable live recovery media built with every release, a guided recovery console that walks the restore, and an automated end-to-end boot test on every change so the media cannot ship broken.
- Storage from retired devices and orphaned backups is reclaimed by the cleanup sweep. Organizations with old retirements should see their backup storage footprint drop.
- Three recovery-console fixes: the reboot countdown no longer hangs, re-authenticating mid-download extends the session correctly, and the recovery download exchange decodes reliably.
Foundations
- Fleet Designer, a new kind of AI agent, reads a bounded snapshot of an organization — devices, software, network, posture, health, automation, event log — and writes an eight-section fleet report as a PDF. It is read-only and changes nothing on the fleet. Alongside it, every device can carry a function classification (AI-derived or set by hand) shown on the device page and usable as a filter.
- Microsoft 365 tenant sync groundwork: schema and sync plumbing for users, Intune devices, conditional access policies, licence SKUs, and Secure Score, plus a manifest update that asks existing customers for four more Graph permissions with a non-interrupting re-consent prompt. Nothing is visible in the UI yet; this is the base for the posture program.
- Software policies can now express 'install this if it is missing' with an explicit arming step and MFA-gated authorization. Nothing dispatches an install yet; this release makes the intent expressible and audited.
- Mobile approver devices are now split into hardware-attested and not attested in the approver list, and a banner walks the technician through registering their phone.
If remote desktop has been dropping on you every minute or so, this is the release that fixes it. Since early August a background job that cleans up abandoned sessions had been ending healthy ones — it only recognised one kind of connection as proof of life, and the default peer-to-peer transport never uses that kind. Sessions were being finalized while a technician was mid-click. The fix lives entirely on the server, so there is nothing to update on the endpoints. While we were in there, Windows hosts with two monitors at different scaling levels now put the cursor where you actually clicked on the second screen; that one does need the updated agent, which arrives on its own.
The biggest new capability is AI script authoring. Ask the assistant to write a script and it drafts a proposal instead of running something. A separate reviewer model — one that never sees the conversation that produced the script — scans and classifies it, and you approve on a card that lays out the goal, the expected effect, the reviewer’s findings, the risk tier, the target devices, and the code. Higher-risk proposals need the scripts permission plus MFA to approve. After a run, Breeze reads the device back independently to confirm the script did what it said before it can be saved to your library. There is an unattended lane for organizations that want reviewed scripts to run without a click, but it stays off until the MSP raises the ceiling and each organization opts in. Alongside it, service deliverables give you a place to track what every contract owes the customer, with key-date reminders, template sets you apply in one step, evidence attached automatically, and a Service page and Documents page customers can see in the portal.
Monitors reshape how you write alerting: one object holds the condition, severity, notification, and response, and Breeze compiles it into the rules and automations that already run, so nothing you have set up changes. Automations move under a Jobs entry, the device list becomes Devices & Assets, and Cmd+K now jumps you back to recent devices. The security pass closes a Stripe checkout link that could stay payable after its invoice was voided, encrypts integration credentials that were sitting in plain text, and enforces site scope on alert correlations. Bare-metal recovery gains Linux live media and a guided console tested by an automated boot on every change. Self-hosters: 22 migrations apply automatically, and if you run a split-role deployment or M365 actions you must set APP_ENCRYPTION_KEY_ID before upgrading — full operator detail is in the GitHub Release notes.