Describe the problem in plain English. eLLA finds it, explains it and fixes it, only with your approval.
eLLA is an AI administration partner that lives in your terminal on Ubuntu and Windows. It investigates with read-only checks, explains what it found, proposes the exact fix with a second AI's review, and only with your approval applies it, verifies it and records how to undo it. If a change cannot be undone, eLLA tells you before you approve. It talks only to your own eLLM endpoint. New: ask "would this machine pass Cyber Essentials?" and get a readiness report against six security standards.
ella undo rolls backA careful second pair of hands for the whole estate
eLLA brings the specialist knowledge and asks the right questions, so a small team can look after Linux and Windows with the same confidence.
IT teams and MSPs
Looking after mixed Linux and Windows estates with too few people. One assistant, one way of working, one audit trail across both.
Generalists
Not a specialist in every platform? eLLA knows the commands, follows proven playbooks, and asks when a decision is yours: which OU, which option, when to reboot.
Experienced admins
A fast, careful second pair of hands that does the digging, documents everything it does, and never runs a change you have not seen.
"Would this machine pass Cyber Essentials?" Now you can just ask.
eLLA checks a device against the security standard you are working towards and produces a client-ready report in seconds: every control, what it found, why it fails and exactly how to fix it. Then it helps you fix it.
Six standards built in, from Cyber Essentials and IASME to the Defence Cyber Certification, ISO 27001, the Essential Eight and PCI DSS. Know where every machine stands before the assessor does.
Cyber Essentials
v3.3 "Danzell", 2026Every technical question: firewalls, secure configuration, security updates, user access and malware protection, plus the Cyber Essentials Plus audit tests.
Defence Cyber Certification
Def Stan 05-138 · Applicant Guide v2.2All 148 controls. Around 55 are checked on the machine itself; the rest are listed as the evidence items you will need to produce.
ISO/IEC 27001:2022
Annex AThe technological controls in A.8, plus endpoint evidence that supports the people and physical controls.
ASD Essential Eight
November 2023All eight mitigation strategies, from application control and patching to macro settings, admin privileges, MFA and backups.
PCI DSS
v4.0.1The requirements that apply to a workstation or server in or connected to the cardholder data environment.
IASME Cyber Assurance
Levels 1 and 2The Cyber Essentials prerequisite plus the IASME additions, for organisations going a step beyond CE.
Checks the machine itself
Firewall, updates and the 14-day patch window, supported OS and software, accounts and admin separation, password and lockout policy, screen lock, anti-malware and web protection, allow-listing, encryption, logging, backup, TLS and more. Read-only, in seconds.
Asks what a device cannot prove
Boundary firewalls, MFA on cloud services, account processes. Answered once per organisation, then reused for every machine you assess.
Clear, client-ready results
PASS, FAIL, EVIDENCE NEEDED or ADVISORY for every control, with what was found, why it fails and exactly how to fix it. Markdown and JSON reports.
Fix it with eLLA
"Help me fix the Cyber Essentials failures" walks through each one, with the usual review, your approval and an undo wherever one is possible.
Ask, investigate, explain, propose, approve, verify
Each session starts with a snapshot of the machine, so eLLA is grounded in what is actually there rather than guessing.
Ask
Type a question or task in plain English, or a quick /command.
Investigate
Read-only diagnostics run automatically: logs, services, disks, network, event logs, AD health.
Explain
A plain-English diagnosis: what is wrong, why, and what it recommends.
Propose
The exact command or file change, the reason, a second-AI review verdict and the undo, or a clear warning that there is none. Multi-step work arrives as a numbered plan.
Approve
Yes, no, or "always for this session". Nothing changes until you say so.
Verify and record
eLLA checks the change worked, logs it, records the undo where there is one and can write an incident report.
you › why did the server reboot last night? read journalctl -b -1 · dmesg · systemctl --failed · free -h eLLA › The kernel's out-of-memory killer fired at 02:14 and the watchdog rebooted the box. php-fpm had 48 workers at ~180 MB each on a 8 GB server during the nightly Moodle cron. Proposed /etc/php/8.3/fpm/pool.d/www.conf pm.max_children 48 → 24, then reload php-fpm Review second AI: safe, reversible, matches the diagnosis Undo backup saved · restore with ella undo Apply? [y/N] y ✔ php-fpm reloaded · 24 workers · memory headroom 2.1 GB · logged
you › create Jo Bloggs in Sales with the same groups as Sam read domain OU layout · username pattern · UPN suffix · Sam's groups eLLA › Your pattern is first initial + surname, so: jbloggs, jo.bloggs@contoso.local, in OU=Users,OU=Sales. Sam is in 6 groups. Plan: 1. New-ADUser jbloggs in OU=Users,OU=Sales 2. Add to the same 6 groups as Sam 3. Require a password change at first sign-in Review second AI: matches request and naming convention Undo remove user and group memberships Initial password: (hidden prompt, never seen by the AI) Approve plan? [y/N] y ✔ 3/3 steps verified · logged to Windows Event Log
An AI that can change servers needs brakes it cannot talk its way past
eLLA's safety does not depend on the model behaving. The rules are deterministic, the review is independent, and everything is recorded.
A classifier, not a judgement call
Read-only commands run instantly; anything that changes state is gated. The AI is never asked to judge safety. On Windows it uses PowerShell's own parser to know exactly what a command will do.
Second-AI review
Every proposed change is reviewed by a second model before it reaches you, and the verdict is shown alongside the command.
Type the machine's name
Domain controller promotion, certificate authorities, disk wipes, firewall shutdown and Default Domain Policy edits need the computer's name typed, even in auto-approve mode.
Undo where it exists, a warning where it does not
Reversible changes are recorded with their undo, and eLLA backs up files, Group Policy objects and registry keys before touching them, so ella undo is an exact restore. Some actions cannot be reversed; eLLA tells you so before you approve, never after.
Never sees your passwords
Passwords go into a hidden prompt and straight to the command. Neither the AI model nor the logs ever see them.
Fully audited
Every action, automatic or approved, goes to an append-only JSONL log plus syslog on Linux or the Windows Event Log.
Private and local-first
eLLA talks only to your own eLLM endpoint. No cloud service is required: run a local model, or route heavier work to a model you choose through eLLM.
You set the pace
--paranoid asks before everything; --yolo auto-approves but still logs every action and still stops for high-impact changes.
Diagnose, audit, plan and build, in one conversation
Health report and audits
One command for security, updates, logs, services, disk, performance, network and time, or focused audits such as /security, /updates, /disk and /net. Clear PASS, WARN and FAIL checklists with prioritised recommendations.
Log watcher
"Watch this service and tell me what happens." eLLA monitors in the background and gives you a full report when you stop it.
Plan-first provisioning
"Turn this server into …" becomes a numbered plan. Approve once, watch it execute with progress, verify each step, and get a written report.
Knowledge that grows
Bundled skills, scripts plus expert playbooks, load automatically when relevant. eLLA can save useful scripts and lessons as new skills, and skills can be shared between installations.
Reports, Git and GitLab
Markdown incident reports with findings, actions, undo references and verification. Git reads run automatically, writes need approval, and eLLA can raise GitLab issues.
Small model by default
Everyday tasks run on small local models such as qwen3-4b. An optional heavy tier, routed through eLLM, is used automatically for plans, policy building and change review.
Bash on Ubuntu. PowerShell on Windows.
eLLA runs the same commands an experienced administrator would, on each platform's own terms.
Linux
Ubuntu 22.04 and 24.04 LTS, desktop and serverHighlights
- bash-native, with
sudoonly where needed - Throwaway test containers in Docker to reproduce an issue or trial a configuration before touching the host
- Remote hosts over SSH: named inventory, key-based access setup, file transfer, approval-gated remote changes and per-host audit
- Config drift baselines to spot unexpected changes
31 bundled skills
Windows
Windows 10 and 11, Server 2019, 2022 and 2025 including Server CoreHighlights
- PowerShell-native, with a safety classifier built on PowerShell's own parser
- Automatic backups before change: Group Policy objects and registry keys captured first, so undo is exact
- Works on modern terminals and the classic console, including Server Core
- An MSI that can itself be deployed by Group Policy or Intune
What it checks
| Event logs | Unexpected restarts, blue screens with stop codes, crashes and hangs, disk and NTFS errors, Group Policy failures, failed logons and lockouts |
| Windows Update | End-of-support status, last patch, pending reboot, WSUS and Update for Business, failures with error-code guidance |
| Services | Stopped automatic services, crash loops, start timeouts, service-account logon failures |
| Storage | Free space, disk health and wear, Storage Spaces, VSS writers, space hogs |
| Network | IP, DHCP, DNS, DC discovery, "trust relationship failed", proxy, firewall profiles, listening ports |
| Performance | CPU, top processes, memory and commit, paging, disk latency, handle leaks |
| Security | Defender, SMBv1 and signing, RDP/NLA, admins, UAC, LSA, Credential Guard, BitLocker, Secure Boot, LAPS |
| Time | Domain hierarchy, PDC emulator, offset against Kerberos limits |
15 bundled skills
Directory and policy work in plain English, for teams that are not Windows specialists
- AD health (
/ad): domain controllers and their support status, FSMO roles, replication, dcdiag, SYSVOL, DC DNS records, functional levels, Recycle Bin, krbtgt age, privileged groups and stale accounts. - Starters and leavers from a sentence: new users copying a colleague's groups, reversible leavers, password resets, unlocks, and finding why an account keeps locking.
- Follows your conventions: learns your OU layout, username pattern and UPN suffix before creating anything, and asks when it is not sure.
- Policy from a catalogue of 60+ vetted settings: removable storage, screen lock, banners, WSUS and update rings, Defender, firewall, RDP, BitLocker, LAPS, SMB/NTLM/LSA hardening, PowerShell logging, browsers, OneDrive, Office macros, power and time.
- Respects the dangerous bits: never edits the Default Domain Policy without a typed confirmation, and steers password and lockout rules to Fine-Grained Password Policies.
- GPO troubleshooting (
/gpo), plus DNS (/dns) zones, scavenging and forwarders, and DHCP (/dhcp) authorisation, scope exhaustion and failover.
lock the screen after 10 minutes for all PCsComputer-scope GPO, linked to the right OU, backed up, verified with gpresult
block USB storage for SalesRemovable-storage policy scoped to the Sales OU, with the scope explained first
point PCs at our WSUS serverWindows Update settings from the vetted catalogue, reachability checked
why does Sam's account keep locking?Lockout source traced through DC event logs, then a proposed fix
Jo has left, sort out her accountDisable, move and remove groups, reversibly, with the undo recorded
Linux and Windows, side by side
| Capability | Linux | Windows |
|---|---|---|
| Conversational diagnosis and fixes | ✓ | ✓ |
| Approval gate and second-AI review | ✓ | ✓ |
| High-impact typed confirmation | — | ✓ |
| Transactional undo, with a warning when none exists | ✓ | ✓ |
| Automatic pre-change backups | files | files, GPOs, registry |
| Hidden password prompts, never seen by the AI | ✓ | ✓ |
| Audit log | JSONL + syslog | JSONL + Event Log |
| Health report and focused audits | ✓ | ✓ |
| Log watcher | journal + logs | Event Log + IIS logs |
| Plan-first provisioning | ✓ | ✓ |
| Bundled skills (save and share your own) | 31 | 15 |
| Local and heavy AI model tiers | ✓ | ✓ |
| Remote hosts | SSH | — |
| Test containers | Docker | — |
| Active Directory, Group Policy, DNS, DHCP | — | ✓ |
| Installer | .deb, offline | .msi, GPO / Intune |
Where eLLA runs today
Ubuntu Linux
22.04 LTS and 24.04 LTS, desktop and server.
Self-contained .deb package, installs offline.Windows desktop
Windows 11, and Windows 10 (supported, flagged as out of Microsoft support since October 2025).
.msi installer, deployable via Group Policy or Intune.Windows Server
2019, 2022 and 2025, Desktop Experience and Server Core.
.msi installer.sudo rights. Windows: an elevated terminal for changes, built-in Windows PowerShell 5.1, and RSAT or role tools for Active Directory, Group Policy, DNS and DHCP features.See eLLA on your own servers
We will connect eLLA to an eLLM endpoint and walk through a real diagnosis on a Linux or Windows machine of your choosing, or try the preview.
