Docs / Get started
What is ivralo
ivralo is an AI SRE for code that cannot leave the network. It runs on your own machine or server, works inside one git repository, talks to a model you choose, and keeps its state in your own PostgreSQL.
Two modes, one engine
ivralo is one engine that runs in two modes. A mode is only a set of grants: what an unattended run may do, and what has to ask a person first.
| Mode | Where it runs | What it does |
|---|---|---|
| Server mode | Next to production | Investigates incidents and reports with evidence |
| Developer mode | In a person's terminal | Reads, edits, checks and commits code, a person approving each edit |
Server mode
In server mode ivralo runs next to your service. The Investigator reads an incident's evidence (the logs, the deploy log, the metrics and the alert itself) beside the service's repository, and files a report in which every claim cites a line of the evidence or a line of the code.
An investigation starts in one of three ways:
ivralo incident, from a terminal.POST /incidenton the local API, whichivralo serveopens on loopback, behind a token.- An Alertmanager or PagerDuty webhook, which becomes a job. The same alert twice is one job.
The Investigator never edits. When a request in server mode does ask for a code change, ivralo can open a draft pull request from a branch it made itself. It never merges and never deploys.
Server mode and incidents goes through it step by step.
Developer mode
Developer mode is the default; Working in a repository covers it in full. You run ivralo inside a repository and a main agent, with specialists when a task needs a fresh context or a narrower grant, reads the code, edits it, runs the project's own checks and commits.
- Every edit is shown as a diff and waits for your yes: once, for the turn, or for the session.
- Shell commands go through an allow, ask and deny policy.
- Several Code Engineers can work at once, each in its own git worktree.
- A run nobody is watching, such as a pipe, a job or a schedule, is granted nothing that would have to ask.
Any model, your network
ivralo works with any model you point it at: a local one through Ollama, your company's own OpenAI-compatible endpoint, or a hosted API. A hosted model is announced at every start, and incident evidence is refused on one unless you allow it with identifiers masked.
Sessions, run history, long-term memory and the knowledge index all live in your PostgreSQL. Nothing leaves the machine unless you send it.
Documents and data
ivralo also reads and writes the files work runs on: Word documents, slide decks, spreadsheets and PDFs, charts, SQL over CSV and Excel files, and text from scans. This side shares the same engine and the same approvals.
Getting it
There is no public install yet. Message us and we will set it up with you.