Urika: Letting AI Agents Run Your Exploratory Analysis

My team and I spend a lot of time on exploratory data analysis. Not the final, polished analysis you put in a paper, but the messy stage before that. Comparing methods and figuring out what questions the data can answer. It’s essential work, but it’s repetitive and time-consuming. You run dozens of variations on similar pipelines and manually document every result, most of which go nowhere.

I built Urika to automate that stage.

What It Does

Urika is a multi-agent scientific analysis platform. You give it a dataset and a research question, and it orchestrates a team of specialised AI agents that systematically explore your data. A planning agent designs the analytical approach, and a task agent writes and executes Python code to implement it. Every method is written from scratch for your dataset and question. An evaluator scores each result against your success criteria and passes the scores to an advisor, which decides what to try next.

A literature agent searches relevant papers you’ve added to the project. When the planning agent identifies a capability gap, a tool builder creates custom analysis and modelling tools on the fly, from domain-specific data loaders to specialised statistical measures. When an experiment finishes, a report agent writes up the findings and a presentation agent builds slides. A project summarizer keeps a running narrative of progress, and a finalizer packages everything with standalone reproducibility scripts.

Why I Built It

I was tired of writing the same analysis boilerplate across projects. Profile the data, check distributions, try a few models, compare results, write it up. Every project follows roughly the same loop, but each one requires fresh code and fresh documentation.

I wanted a system where I could describe what I was investigating and have the exploration happen systematically. It would log every approach it tried, including ones I might not have considered, and rank them on a leaderboard so I could focus on the results that best met my success criteria. I chose a multi-agent architecture because, after more than six months of building and working with agentic systems, I could see that the stages of my AI-assisted research workflow (planning, execution, evaluation, advising) mapped cleanly onto specialised agents.

Handling Sensitive Data Responsibly

If you work in academia or with industry partners, you deal with ethical constraints around data constantly. There’s always something that shouldn’t be sent to a cloud API, whether it’s student submissions, unpublished research, ethics-approved participant data, or industry datasets under NDA. Ethics board requirements and data-sharing agreements can restrict where that data is processed. So can consent forms that never anticipated cloud AI, since participants were not asked to agree to it.

Urika has three privacy modes, and hybrid mode is designed for data like this. In hybrid, a private Data Agent reads your raw data using a model inside your data governance boundary, whether that model runs locally via Ollama, on a secure institutional server, or on your organisation’s own cloud infrastructure. It passes sanitised summaries to the agents doing the conceptual work (planning experiments, evaluating results, advising on next steps), which can use frontier cloud models where you need the extra capability. Those cloud models never see the raw data.

You can also run fully private (everything on local or institutional endpoints) or fully open (everything cloud) depending on the project. The routing is per-agent, so you decide which models handle which parts of the workflow.

Two Ways to Work

In guided mode, Urika runs one experiment at a time. You read the result and the evaluator’s assessment, then decide what to try next. Between experiments you can also chat with the advisor, which sees every method and result in the project when it makes a suggestion.

In autonomous mode, the advisor agent takes over. It reviews results and launches follow-up experiments on the gaps it finds, without waiting for you. You come back to a complete project book with a leaderboard of every method attempted. Urika also connects to email, Slack, and Telegram, so you can get progress updates and even pause, stop, or query a run remotely.

Urika ships with three first-class interfaces that share the same project state on disk:

A scriptable CLI for batch jobs and CI. Every command supports --json output for custom pipelines.
An interactive TUI for exploratory work, with orchestrator chat and tab-completed slash commands.
A browser dashboard for monitoring long runs, sharing results, and managing settings.

A run you start from the CLI shows up in the dashboard on the next refresh.

Where It Is Now

Urika is open source (Apache 2.0) and in early, active development. It runs on Linux, macOS, and Windows 11. Urika is pip-installable and works with multiple model providers. Open and hybrid modes need at least one cloud provider configured (Anthropic, OpenAI, or Google). Fully private mode runs on a local or institutional endpoint and needs no cloud account. The Anthropic adapter can authenticate with either an API key or a Pro/Max subscription, and the Claude Code CLI is optional. If you authenticate with an API key, set a monthly spend limit in the provider’s console before you start.

More adapters are on the roadmap, including one for PI. I want Urika to run on the models and servers your lab already has.

If you work with data and want to speed up the exploratory phase, give it a try. I’d appreciate feedback: bug reports, feature requests, or just hearing how it works on your data.

GitHub: github.com/xkiwilabs/Urika

Urika is part of the xKiwiLabs open-source toolkit. Everything is free and built in the open for researchers and practitioners who want AI tools that fit their workflow.

Michael Richardson Professor, School of Psychological Sciences Faculty of Medicine, Health and Human Sciences Macquarie University

AI Disclosure: I wrote this article with help from AI tools, including Claude. The ideas and the workflow described here are my own. The AI helped me draft and edit the text. I use AI openly in my research and teaching, and I encourage others to do the same.