OmniScientist V2 CLI (omni)

A local-first, CLI-first personal research agent. omni plans work, calls skills (Claude Code / Codex / OpenClaw SKILL.md compatible), runs a bounded ReAct loop, remembers across sessions (SQLite + filesystem only), and ships with WeChat / Feishu / DingTalk channels.

This CLI is the official next-generation implementation of the OmniScientist framework described in Shao et al., OmniScientist: Toward a Co-evolving Ecosystem of Human and AI Scientists. The current checkout reports package version 2.0.0rc2 (planned Git tag v2.0.0rc2).

This folder (cli/) is the Python application. The research skills are a separate, independent top-level collection in the repository's skills/ directory. The CLI runs skills; the skill packages do not depend on CLI internals.

New users should start with the repository-level installation and initialization guide. This file is the package/developer quick reference.

Install

omni runs on macOS, Linux, and Windows. It requires Python ≥ 3.11 and Node.js ≥ 20.9 with npm for the bundled research-pptx renderer. PyPI is the stable package authority; install into one isolated tool environment:

uv tool install OmniScientist-V2
# or
pip install OmniScientist-V2
# or
pipx install OmniScientist-V2

omni

The first bare omni opens setup, prepares the lockfile-pinned Node runtime, and starts or repairs the single Home Service. Normal updates use one product command:

omni update

omni update delegates package replacement to the manager that owns the running CLI (uv tool, pipx, or an explicit dedicated Python environment), then performs runtime preparation, legacy daemon cleanup, Home Service launcher refresh/restoration, and readiness verification as one serialized lifecycle transaction. The public command is intentionally parameterless; omni update status is read-only diagnostics.

Direct manager updates are supported too:

uv tool upgrade OmniScientist-V2
pipx upgrade OmniScientist-V2

Those commands replace Python packages only. The next bare omni detects the changed package fingerprint and automatically completes the missing local runtime/service convergence before entering the REPL; there is no second update command to remember.

From a downloaded source checkout, use the repository installer. It defaults to an isolated uv tool and deploys the current local snapshot; it never silently selects an active Conda/venv:

./cli/scripts/install.sh
./cli/scripts/install.sh --editable --local  # contributor install

powershell -ExecutionPolicy Bypass -File cli\scripts\install.ps1

Rerun the repository installer to redeploy the current tree. omni update on a linked clean checkout safely fast-forwards its configured upstream and reinstalls. A standalone copy of the installer uses PyPI; the moving master channel is available only as an explicit development choice. Installing into an active environment requires the explicit advanced --method env (-Method env) option.

Then, identically on every OS:

omni --version
omni init                           # model, retrieval mode, workspace, optional exports/MCP
omni doctor                         # environment and configuration checks
omni                                # interactive REPL
omni "Explain diffusion models in one sentence"  # one-shot
omni update                         # owner-aware package + lifecycle update

Like Codex, interactive startup reads a local update cache and offers a one-key update when a newer PyPI release is known. Network refresh happens in the background (TTL 24h), stays silent offline, and can be disabled with omni config set update.check false or OMNI_UPDATE_CHECK=0.

Full per-OS install/update/serve walkthrough: getting-started.md; platform table: commands.md.

First-run model setup

omni init walks you through it (provider menu, default openai). omni otherwise defaults to an offline mock model so it runs with zero config. Point it at a real provider in one line (works on the shell and inside the REPL, where model changes take effect immediately):

omni config model -p deepseek -u https://api.deepseek.com/v1 \
  -m deepseek-chat -k "$DEEPSEEK_API_KEY" --test
omni config list                              # provider / model / endpoint / key status

-p accepts openai / deepseek / ollama (all OpenAI-compatible; base_url picks the real service) or mock. Prefer files? Edit ~/.omni/config.toml ([model]) and ~/.omni/secrets.toml ([model] api_key). The key is stored in secrets.toml and masked everywhere.

Shell and REPL commands

Use omni <command> ... in a terminal or start omni and use /<command> ... in the REPL. Both forms dispatch to the same command implementation and configuration:

Shell                         REPL
omni config list              /config list
omni task list               /task list
omni channel list             /channel list
omni serve status             /serve status
omni autosota status          /autosota status

In a capable TTY the REPL commits its transcript to the terminal's native scrollback above a bottom dock containing the composer and permanent status line. It stays in the normal buffer with no mouse capture, so use the terminal's own selection (click-drag, or Shift-drag/copy-mode in a multiplexer) to copy history and its usual wheel/keys to scroll. Blank Enter is ignored. auto falls back to the classic prompt for pipes, redirection, CI, TERM=dumb, and unsupported terminals. Enter sends; Ctrl+O and Alt/Option+Enter insert a newline, as does Shift+Enter when the terminal reports the modifier. Use trailing \ + Enter as a portable continuation, or Ctrl+X Ctrl+E to edit the draft with $VISUAL/$EDITOR. Multiline paste remains one request. Override the UI mode with omni --ui tui|classic, OMNI_UI=tui|classic, or omni config set display.ui_mode auto|tui|classic (tui selects the inline dock). Committed transcript text stays in shell scrollback after exit, and Omni's conversations and tasks remain persisted.

Natural language has no slash inside the REPL. Channel messages are processed by the persistent omni serve owner, not by an open REPL window. See the complete channel lifecycle and login guide. Slash commands do not invoke a shell, so $ENV, pipes, and redirections are not expanded; omit secret options to use a hidden prompt.

Skills

Every skills operation has both a terminal form and a slash form after starting omni:

Local paths may be a skill directory, SKILL.md, or another .md file; tool:name supports claude:, codex:, agents:, and openclaw:. Direct Git imports use the default branch. See skills.md for the complete source matrix and limits, installation, trust, natural-language and explicit invocation, loading, and cross-agent comparison. See skills/README.md for the built-in catalogue and skills/docs/authoring.md for authoring.

Research commands

omni is a research agent, not a coding one: it grounds answers in a local literature corpus (omni lit, cited [S#]), records an auditable hypothesis→claim→evidence trail (omni hypo/claim/evidence), keeps a reproducible run ledger (omni run), and audits its own claims (omni verify) / retrieval (omni bench). All are also REPL slash commands. See commands.md and research-agent-design.md.

Docs

• getting-started.md — install, configure, first run, research workflow
• commands.md — command reference (incl. research commands)
• skills.md — skill install, discovery, automatic/explicit invocation, and agent comparison
• testing-and-evaluation.md — native test stack and Inspect AI selection
• research-agent-design.md — design, capabilities & how it compares
• architecture.md — request flow & internals
• compatibility.md — Claude Code / Codex / OpenClaw / MCP

Develop

cd cli
uv run --extra dev --extra mcp --extra vec --extra channels pytest -q
uv run --extra dev ruff check src
uv run --extra dev omni eval --coverage
uv run --extra dev omni eval --research-quality

Apache-2.0. See LICENSE, SECURITY.md, and PRIVACY.md.