Software that
evolves itself.
Ops AI Bot is a product in which AI does the operations and maintenance of a system. From development and testing to production, the AI manages issues, fixes them, remembers and reports. People only set the direction and approve.
grows with use
* 2 minutes and 0 operations were measured in a development environment on 12 November 2026.
The worry in maintenance is depending on people.
People leave, fall ill or take holidays; knowledge stays in one head; incidents happen at night; reports have to be written. Ops AI Bot removes this dependence on people with a mechanism.
BEFORE ── maintenance done by people
- Someone is on call at night and on holidays
- They read logs, fix, and release by the runbook
- Knowledge lives in their head and leaves with them
- People write reports and hand-over documents
AFTER ── maintenance handed to Ops AI Bot
- The AI watches 24 hours, detects and opens the ticket itself
- The AI investigates, fixes, tests and opens the PR
- Causes, fixes and decisions are remembered by the system
- Plan, evidence and result stay in the ticket — no separate report
AI does the "work" of operations and maintenance.
Manages its own tickets
Opens tickets from chat posts, errors and CI failures. Numbers, states, history, evidence and PRs are all kept.
Talks in your team chat
Picks up requests and questions even without @, reads images, PDFs and logs, and replies with the ticket number and result.
Plan first, fix after OK
For change requests it shows a plan, waits for OK, writes tests, fixes, and opens a PR on its own branch.
Automatic release in development
Only changes with all CI green are swapped in via the approval gate. On failure it rolls back to the previous version.
Production needs a human
The AI only proposes. The approval gate checks the approver, proposer ≠ approver and the plan before executing.
Memory prevents gaps
Remembers causes and fixes, procedures and team decisions for the next judgement. Never personal data.
Just write, and the AI acts.
During development and testing
Product testers, designers, developers and QA simply write in the chat group (LINE, Teams, Slack) where Ops AI Bot is. No @ needed. The AI turns it into a ticket, shows a plan, fixes and tests after OK, releases to development and reports. It also remembers decisions and items on hold.
After go-live
The group is your owner, your maintenance staff and Ops AI Bot. 24/365 the AI watches, detects, analyses, fixes and verifies; the only human job is to approve — one tap on a phone. Improvement requests after go-live follow the same flow.
What differs from ticket tools is
who does the work.
Ticket tools record and visualise the work people do. In Ops AI Bot the AI does the work itself. Neither is better — they cover different ground.
| Traditional operations tools | Ops AI Bot | |
|---|---|---|
| Who does the work | People (AI assists with summaries and reply drafts) | AI (people set direction, judge and approve) |
| Noticing | A person watches the alerts | The AI detects and opens a ticket |
| Investigating and fixing | Staff read logs and fix | The AI investigates, fixes, tests and opens a PR |
| Releasing | A person follows the runbook | Automatic in development; the approval gate executes production |
| Reporting | A person writes the report | Plan, evidence and result stay in the ticket automatically |
When an incident happens at night
The AI detects it and opens a ticket (nobody is woken up)
It investigates, fixes and tests, and prepares a PR
In the morning, staff check it on their phone and approve
It is released by the fixed procedure; the report is kept automatically
An AI that decides, an AI that works.
Ops AI Bot is an independent platform on a server separate from the business system. From the target it only reads logs, health, repository and CI status. It writes only PRs to Git and approved releases.
Six entry points
Chat posts (with images, PDFs and logs), error logs, health checks, CI results, Git events and schedules. Human requests are always accepted; work the AI infers is handled only when it matches the team's rules.
Router and Worker
A Router stays per target and accepts, routes and answers (no shell, no file writes). A Worker starts per ticket and investigates, fixes, tests and opens PRs in an isolated workspace.
A program decides the change class
A program — not the AI — classifies a change from its paths and content. Changes to money, permissions, personal data or DB schema need extra checks. The AI cannot loosen its own rules.
Only the approval gate holds production keys
The approval gate is a program that never calls an AI. It checks the approver, proposer ≠ approver, that plan and verified build match, and freeze periods before it executes.
Rolls back by itself on failure
If health does not return after a development release, it rolls back to the previous version and tells people and the AI. The target server accepts only fixed commands (release, rollback, restart).
Choose your model
Connects through OpenAI- or Anthropic-compatible APIs, including models you run in-house. Router and Worker can use different models.
Widen what it may do, step by step.
| Level | What the AI does |
|---|---|
| L0 Observe | Records its judgement only |
| L1 Analyse | Analyses causes and reports |
| L2 Fix proposal | Opens a PR (a person merges) |
| L3 Dev automatic | Merges, releases and rolls back in development |
| L4 Verify and propose | Verifies in staging and proposes production |
A level that changes production on its own does not exist.
Even if manipulated, it cannot get out.
Separate from the business
A separate platform from your business system: no shared servers, networks or secrets.
The AI holds no secrets
No keys or tokens in the AI's sandbox, and no outbound network.
Everything is recorded
Judgements, work, approvals and executions go into a tamper-evident audit log.
Permissions that expire in minutes
Git permissions are per repository, issued each time as tokens that expire in minutes. The core, not the AI, pushes and opens PRs.
Ignores smuggled instructions
Instructions inside logs or messages are treated as untrusted data. Rules and permissions cannot be changed by text.
Stop it any time
Pause everything, freeze for a period, lower the autonomy level, or stop a single ticket.
What you need
| Platform server | 2 vCPU, 8 GiB memory, 64 GB disk or more (add more for many parallel Workers). Linux (verified on Ubuntu 24.04 LTS) + Docker Compose v2 |
| Where | One dedicated server in your cloud or data centre. Verified on Azure; the same setup runs on other clouds and on-premises |
| Network | Inbound 443 and 80. Outbound only to the AI model, GitHub and chat APIs (through an allow-list proxy) |
| AI model | An OpenAI- or Anthropic-compatible API (contract and API key). Models you run in-house work too |
| Integrations | GitHub (repositories, CI) and the chat for the operations group (LINE, Teams, Slack) |
| Target system | Start with a health-check URL and read access to logs. No change to business logic (settings must be added) |
How to adopt ── start by observing, widen as you verify
Assessment: check target, environment, CI and release method; list the settings needed
Observe: connect and check that the AI's judgement matches people's
Fix proposals: connect chat and Git, check PR quality
Automate development: release and roll back CI-green changes automatically
Production approval: release plans executed by the approval gate
Availability (November 2026)
Available now
- Requests, questions and approvals in chat
- From ticket management to PR
- Automatic release and rollback in development
- Approval gate, change classes, autonomy levels
- Console, memory, audit log
Coming next
- Release plans for staging and production
- Automatic daily summaries
- GitLab, Jira and other integrations
- Several systems on one platform
- SaaS offering
AI employees work with your staff;
Ops AI Bot grows your system.
Start with one system, by observing. The assessment lists the settings and the work needed.