Skip to content
Arboria

Arboria

An AI-native ML research workflow that wraps the compute layer. State lives in a ledger, the notebook is yours, Claude writes the code and commands, and you approve every cluster run in your own terminal.

Quick start (Chinese)GitHub
idea 树：每个 idea 一行，从某个 run 分叉，竖条是 run

Seeds

From a seed to an experiment

An idea with no repo yet is a seed, with its own marimo notebook in ~/arboria. When it is ready, arb seed plant turns it into an idea in a repo — a branch, a worktree and a notebook that starts with the seed’s notes.

Seeds (Chinese) →
The Arboria seed garden: four seed cards (thinking, planted) and one todo waiting for Claude
The seed garden (arb ui outside a repo): thinking, planted, dropped, and todos for Claude.
The marimo notebook of the seed loss-spikes: the idea, hypothesis and metric, then observations and a guess
A seed notebook: observations, a guess, a toy experiment. You write it; Claude only reads.

Notebook

The notebook is the bench

Each idea has one notebook, and only you write it. lab.run creates a review card; results and curves come back to the same page. Analysis commands are written by Claude once, saved as functions, and then run without spending tokens.

Notebook and lab (Chinese) →
lab.plot_loss in an idea notebook comparing the loss curves of two runs; the baseline has spikes in the first 2k steps
lab.plot_loss is a project command Claude wrote; the line below names it and its version.
Output of lab.run: r003 awaiting approval, with the review card expanded: checks, resources, estimate, budget and card hash
lab.run only creates a review card: snapshot, resources, estimate, budget, hash. Approval happens in your terminal.

Dashboard

Everything at a glance

arb ui opens a read-only dashboard: what is waiting for you, how ideas branched over time, and where the compute went. No buttons that act; actions are shown as commands to copy.

Dashboard (Chinese) →
The Arboria dashboard: counts on top, what needs you on the left, the idea tree and idea details on the right
Left: what needs you. Right: the idea tree and the selected idea.

Approval

Only you approve

Claude and the notebook can only create review cards. A job reaches Slurm only from your own terminal, after you type the run id; if anything on the card changes, the old approval no longer counts.

Submitting and approving (Chinese) →
Only you approve: the path from a notebook to the clusterClaude Codecode and commandstodos ⇄ codecannot approveNotebooklab.run(…)Review cardsnapshot · resources · hashYour terminalarb approve r014Slurmsbatch, segmentslocal onlyarb watch: state, SU, ARBORIA_OUT pulled back

Why

  • State lives outside the chat

    Ideas, runs, notes and decisions live in a local SQLite ledger that only arb writes. A new session gets a one-screen summary from a hook.

  • The notebook is yours

    Each idea has its own branch, worktree and marimo notebook. Claude can read it, not edit it; its work comes from the todos you leave.

  • Claude writes code and commands

    Write lab.svd(r, "plot the top 20 singular values") in the notebook. If the command does not exist, Claude writes it and the cell re-runs.

  • The compute layer is wrapped

    Segments, billing (with DeltaAI’s rules), budgets, Duo logins and early cancellation are handled by arb.

  • You approve every run

    The review card is bound to a hash of snapshot, command, resources and environment. No code path approves on its own.

  • Guard rails, not a sandbox

    They stop Claude from going along with a stray instruction, not a deliberate adversary running as you on the same machine.

Try it locally first

Install arb and create an idea and its notebook in any git repository. Before running on a cluster you need a machine profile. The documentation is in Chinese.

Quick start (Chinese)Concepts (Chinese)