Senior-level command of your system,
for the whole team.
The on-premises AI harness for your collection system. RAIT enriches the assistant you already use with context from your configuration, code, documentation, and live data: answers in minutes with sources, concrete grounded fixes, and every change reviewed and reversible.
Get a DemoGrounded in your system
Answers come from your real configuration, code, and manuals, with the sources they used.
On-premises and private
Runs on your hardware, on open-source foundations. Your data stays inside your environment.
Works with any AI
Claude, ChatGPT, Gemini, Microsoft Copilot, and others all connect to RAIT the same way.
Every change reviewed
Configuration and code are version-controlled. Nothing changes without review, attribution, and a reversible history.
One question turns into an investigation. Then into a fix.
You work in the AI assistant your team already uses. RAIT harnesses it, enriching every question with context from your configuration, your code, your documentation, your live data and your own change history. Pick a scenario.
Four pillars
Three are available today. The fourth is on the roadmap and will be included in the base product.
Insight
Available nowAsk "how does this actually work?" in plain English and get an answer grounded in your system, with the sources it used. Hours of digging become minutes.
Control
Available nowConfiguration and code are versioned in one place. Drift is detected, changes are shown as a clear before-and-after, and every edit is reviewed, attributed, and reversible.
Advice
Available nowBring a discrete problem and get a concrete, grounded proposal you can act on: the likely cause and a specific, reviewable fix.
Creation
On the roadmapThe AI drafts the configuration change or routine for you, then routes it through the same review and gating before anything lands.
Capabilities
The specifics behind each capability, for the people who will run it and the people who have to approve it.
Knowledgebase
Semantic and keyword search over documentation, product manuals, configuration, and code. Office files and PDFs are parsed on upload, with text, tables, and figures preserved and searchable. Embeddings are generated on your server by an open-source model, so nothing leaves your network. Learnings from your own team are reviewed, attributed, and retained alongside the documents, and an approved learning outranks vendor text when the two conflict. Every answer cites the document, section, and object it came from.
Change control
Detects drift between the running system and the versioned copy of it, whoever made the change and by whatever route. A semantic diff shows each change in product terms rather than raw storage internals, laid out like the system's own configuration screen. Each change is attributed with a reason, a ticket reference, and a reviewer. Peer review runs as a pull request, and approval gates the deploy. By default no one approves their own change, and each repository can run in peer, solo, or trusted review. On request, your assistant reviews a change that is waiting and drafts comments for a person to post.
The interface
A set of tools your AI assistant uses to reach your estate: search over code, structured queries over configuration, semantic search over knowledge, read-only queries against the live system, and change review, all exposed through MCP. Nothing to install on user machines beyond a small configuration your administrator pushes, and each person signs in as themselves, with no API keys to mint or rotate.
One console
A single browser home for everyone: the Knowledgebase, Change Control, Learnings, and user and role administration. Nothing to install for analysts and reviewers. A fleet board shows the health and version of every service at a glance.
Access and audit
One sign-on through your identity provider, with no separate password store. Directory groups decide which systems and environments a person can reach, and roles with per-person allow and deny rules refine that access. A request is permitted only when some rule allows it and no rule denies it. Unknown, inactive, or errored users are denied by default, and every call, including every denial, is written to an audit trail.
Live system access
Ask questions of the live system safely and read-only, so you can ground an answer or verify a change against what is actually running. It reads InterSystems Caché and IRIS, SQL Server, and MySQL systems. A guard reads each query the way the database engine would and refuses anything that is not a read, and results are capped by rows, cell size, and run time so a query cannot burden production. Scoped per subsystem, and available as an optional module.
Privacy and masking
Consumer PII and PCI data are masked in the tool layer before they reach the model, under a policy you review. Card data, passwords, tokens, and keys are always withheld. Masking happens in the response path, so your data is never altered, and if the policy is missing or invalid, results are withheld rather than passed through.
Deployment and operations
One Linux host running a Docker container stack, with a GPU optional. A checksummed offline bundle and one installer script install RAIT, and a scripted upgrade carries your settings forward and backs up the version-control store before the new release is installed. A signed license file is verified offline, with no license server and no phone-home.
Works with the AI you already use
RAIT is an AI harness. It does not replace your assistant or lock you into one AI vendor; it enriches the one you already use. Everything it offers is exposed through the Model Context Protocol (MCP), the open standard every major assistant now supports. Your team keeps using the assistant they know while RAIT enriches it with the context of your system.
Repository lens
Literal and pattern search across every file in the versioned estate, plus file reads and scoped exports.
Configuration lens
Structured queries over configuration objects: count, compare, and cross-reference them across a subsystem.
Knowledge lens
Semantic search over manuals, documentation, figures, and the tribal knowledge your team has retained, returning the passage and the citation behind it.
Live query lens
Read-only queries against the running platform, scoped per subsystem, to ground or verify an answer. Consumer PII is masked out of the rows before they reach the model.
Change review lens
List drift, assemble a changeset, comment, and request changes from your assistant, through the same workflow as the Console. An automated agent can never approve or merge.
Connect with any MCP-compatible assistant:
Open source, on purpose
| No copyleft | Permissive licenses only, by policy. Nothing obliges you to publish anything you build. |
| No lock-in | No proprietary database or search engine underneath. If we part ways, your estate is plain git and plain files you already hold. |
| Inspectable | Your security team can read the source of every component you run. |
| Self-contained | Installs and runs fully offline. The only outside service is the AI backend you chose. |
Supported collection systems
RAIT is built to serve more than one collection platform. The shared platform, meaning indexing, search, review, identity, audit, and deployment, is reused for every system, so each new one starts from a working product rather than a blank page. What a system needs of its own is real engineering: an extract adapter, a schema mapping, and the system-aware handling that lets its objects be read, diffed, and reviewed properly.
FACS
Production-readyFinvi (Ontario Systems) FACS on InterSystems Caché or IRIS. Full support across Insight, Control, and Advice, grounded in FACS configuration, ObjectScript code, and the FACS product manual. The estate can be extracted through a read-only SQL account, with no operating-system login on the FACS server, and changes are drawn as the FACS windows they belong to.
Artiva
Production-readyFinvi Artiva RM and Artiva HCx, which are the same target from RAIT's point of view. Configuration and Artiva BASIC logic-block scripts are extracted through a read-only SQL account, with nothing installed on the Artiva server and nothing written to it. They are versioned and reviewed like any other change.
SQL Server and MySQL systems
Production-readySystem Modules for collection systems that run on SQL Server and on MySQL bring live query and change control to them. Tables, views, procedures, functions, and triggers are put under the same version control, drift detection, and review as the rest of the estate.
TCS / CUBS
In progressOnboarding is underway on the shared platform: extract, schema mapping, and system-aware handling are in development.
Your system isn't listed?
Contact Rob to discuss the plausibility of extending RAIT to support your platform.
Secure and private by design
On-premises
The whole platform runs on one host inside your network, on Docker. Read-only first, with live changes gated behind review.
Open-source foundation
Built on permissively licensed open source for storage, search, and version control. No proprietary lock-in underneath.
Your data stays yours
The only AI RAIT runs itself is a self-hosted, open-source search model. RAIT sends nothing to an outside service.
Private cloud path
When you use a cloud-hosted model, RAIT supports a private connection into your own cloud model service, with no public internet in the loop.
Identity, audit, and masking
Single sign-on, role-scoped access, and a full audit trail. Consumer PII is masked before it reaches the model, under a policy you review. Unknown users get nothing.
Reviewed and reversible
Every configuration and code change is attributed, reviewable, and reversible through git-native history.
Scoped per system
Access is granted per product and per subsystem. Someone cleared for one system sees nothing of another, and anything not explicitly granted is denied.
Nothing calls home
There is no license server to reach and no phone-home. Installs and updates arrive as an offline bundle you apply yourself.
You can walk away
The estate RAIT maintains is plain files in git: configuration as JSON, code as source. That history stays readable if you ever stop using RAIT.