Alice CLI

Alice CLI is a governed research and model-development agent maintained by AstralSeek Research.

It combines a terminal agent, tools, skills, persistent sessions, optional messaging, and auditable model-development workflows. Private data and artifacts remain local unless the user explicitly approves an operation that requires another boundary.

Install

Clone the maintained repository directly:

git clone https://github.com/zdqsgithub/alice-cli.git
cd alice-cli

The managed installers provision Alice CLI and its pinned dependencies. The default data home is ~/.alice on Unix-like systems and %LOCALAPPDATA%\alice on native Windows.

Quick Install

Linux, macOS, WSL2, and Termux:

curl -fsSL https://raw.githubusercontent.com/zdqsgithub/alice-cli/main/scripts/install.sh | bash
alice # start chatting!

Windows PowerShell:

iex (irm https://raw.githubusercontent.com/zdqsgithub/alice-cli/main/scripts/install.ps1)
alice

npm terminal CLI (Node.js 20+ and uv required; it provisions CPython 3.12):

brew install uv # macOS example; see https://docs.astral.sh/uv/getting-started/installation/
npm install --global @zdqsgithub/alice-cli
alice --help

See the npm launcher documentation for runtime, diagnosis, and uninstall details. The npm package provides the terminal command; the macOS desktop DMG remains a separate GUI distribution.

First run

Configure a provider, select a model, review tools, and verify the installation:

alice setup
alice model
alice tools
alice doctor

Update a managed checkout with alice update.

CLI and TUI

Start the classic interactive CLI with alice. Start the React and Ink terminal interface with:

alice --tui

Useful commands include:

alice model      # select a configured provider and model
alice tools      # review enabled tools
alice doctor     # diagnose installation and runtime issues
alice update     # update the maintained checkout
alice --help     # inspect the finite public command surface

Inside a session, /help shows available slash commands. Account, subscription, payment, and vendor-portal routes are not part of the Alice distribution.

Governed model-development capabilities

Alice separates intake, development, promotion, and formal evaluation. Plans freeze the target, split policy, primary metric, optimization direction, tie-break rules, candidate space, and budgets before model comparison. Search cannot inspect a sealed test, retains negative attempts, and cannot promote a candidate or open formal evaluation without a separate human approval.

Current bounded capabilities include:

• alice_ml: model-ready CSV/Parquet tables and labeled NPZ matrices for governed binary classification and regression, deterministic controls, immutable artifacts, and aggregate reports.
• alice_research: offline or explicitly authorized evidence workflows.
• alice_autoresearch: optional finite candidate search with a declared search space, append-only ledger, and human promotion.
• alice_graph: privacy-projected advisory graphs over allowlisted aggregate records. Graph values are never model features.
• alice_methylation: the fixed M8 public methylation bundle and external-benchmark lifecycle.

The completed M6 boundary is intentionally narrow. M6A uses the optional cardio_benchmark path for local-only, aggregate-only replay and produces a combined benchmark receipt. M6B supports model-ready multimodal comparison only from one prejoined CSV/Parquet matrix and can compare late_fusion_logistic with declared single-modality baselines. It does not parse, join, embed, or train directly from raw text, PDFs, images, ECG waveforms, audio, or video.

M6.1 remains development-only. M6.1A applies the generic adapter-development contract to a fixed tabular cardio profile with a patient-purged 2020–2023 train and 2024 validation split; it does not reopen the consumed final test. M6.1B proves the same generic path on scikit-learn's bundled public Wisconsin breast-cancer dataset without importing cardio code. M6.1C records the four-state CEO risk audit, where engineering evidence can close or mitigate scoped risks while real-world value remains blocked_external until approved external evidence exists.

M8B performs governed grouped development and a separately approved GSE79775 external evaluation through evaluate_formal exactly once. Install zdqsgithub-alice-cli[alice-ml] and use alice-methylation-m8b-verify for record-only verification without re-evaluation.

M8C is a restricted aging QA path. alice-aging-m8c-verify keeps training_allowed=false, uses synthetic conformance cases, does not read participant media, does not call the aging training launcher, and remote egress remains disabled.

M8D uses alice-m8-verify to derive content-addressed dcau_ and mcau_ completion records from existing evidence. It does not reopen the formal test. The external objective not_met remains the scientific result, and engineering completion is not arbitrary-dataset support.

Privacy and approval boundaries

• Source rows, protected content, credentials, prompts, predictions, embeddings, and model artifacts stay outside aggregate reports and privacy-projected graphs.
• Raw clinical text, PDFs, images, waveforms, audio, and video are not accepted by the generic trainer.
• A project adapter may validate model-ready local data, but it cannot bypass approval, leakage controls, promotion, or sealed evaluation.
• The system does not claim clinical validity, regulatory compliance, deployment readiness, autonomous approval, or a global optimum.
• A successful engineering gate does not convert an unmet scientific objective into a positive result.

Development

Use an external virtual environment and the canonical zero-retry test runner:

git clone https://github.com/zdqsgithub/alice-cli.git
cd alice-cli
curl -LsSf https://astral.sh/uv/install.sh | sh
uv venv ~/.alice/venvs/alice-dev --python 3.11
source ~/.alice/venvs/alice-dev/bin/activate
uv pip install -e ".[all,dev]"
scripts/run_alice_tests.sh --file-retries 0

Operational contracts and milestone-specific boundaries live under docs/alice.

Contributing

For a managed contributor checkout:

curl -fsSL https://raw.githubusercontent.com/zdqsgithub/alice-cli/main/scripts/install.sh | bash
cd "${ALICE_HOME:-$HOME/.alice}/alice-cli"
uv pip install -e ".[all,dev]"
scripts/run_alice_tests.sh --file-retries 0

Run alice update to update that managed checkout. Issues and pull requests belong at zdqsgithub/alice-cli.

License

Alice CLI is available under the MIT License.