Clio Agent

⚠️ Beta. CLIO is under active development — interfaces, blueprints, and on-disk formats may change between releases. Pin a version for reproducibility.

What is CLIO Agent?

An autonomous agent for scientific data management — HDF5, Parquet, and CSV inspection through a FastMCP tool gateway, with a Bubbletea terminal UI (gact) for interactive work. CLIO is the Intelligence Layer (CEI) of the IOWarp platform.

The agent runs in three tiers: a main orchestrator (planner loop over registered tools and experts), domain experts (data / analysis / visualization), and ephemeral nanoagents for parallel sub-tasks. State lives in a local-first memory layer (ARC) that all tiers read and write.

• Multi-expert orchestration with capability-based routing
• FastMCP tool gateway, HDF5 + Parquet servers out of the box
• ARC memory with O(log N) context retrieval
• Optimizer layer for offline tuning + online learning
• Runtime doctor for LM / gateway / file-policy health
• Works with LM Studio, Ollama, OpenAI, Anthropic, OpenRouter, OpenAI Codex subscriptions, and the Argonne ALCF inference gateway

Quick Start

One line installs clio-agent from PyPI, downloads the matching CLIO-branded TUI binary for your platform, and drops a clio command on your PATH.

Linux / macOS

curl -fsSL https://raw.githubusercontent.com/iowarp/clio-agent/main/install/install.sh | bash
clio

Windows (PowerShell)

irm https://raw.githubusercontent.com/iowarp/clio-agent/main/install/install.ps1 | iex
clio

clio boots the server (if it isn't already up) and attaches the TUI. On first connect, pick an LM provider in the modal and you're chatting.

Prerequisites for the default release install: uv or Python 3.12+ with pip. git and Go are only needed when you opt into source-build mode with CLIO_REF or GACT_REF.

CLIO pins its intentional DSPy 3.3 prerelease and the tested stable LiteLLM release. Registry-backed uv tool install needs that exact DSPy dependency declared as an explicit root: use uv tool install --with dspy==3.3.0b1 clio-agent==0.8.0 for a persistent backend-only install. This keeps unrelated dependencies on stable releases. Reserve uvx / uv tool run for disposable checks. See install/README.md for the full contract.

More ways to get CLIO

The install script above is one of several download options:

• 🌐 CLIO website → — guided downloads and a tour of the product.
• 📦 GitHub Releases → — desktop apps (.msi / .exe for Windows, .dmg for macOS, .deb / .AppImage / .rpm for Linux), standalone clio-tui-<os>-<arch> binaries, and the web bundle.
• 🐳 Docker (no install) — docker run -it --rm ghcr.io/iowarp/clio-tui:latest, or docker compose up clio-web for the browser UI.
• 🖥️ Browser UI — clio --web serves the web app locally.

Full matrix (script · docker · web · desktop · compose · source) with copy-paste commands: docs/INSTALL.md.

Using the clio command

clio with no arguments is the one-command UX. The subcommands manage the backing server without hunting PIDs:

Override paths with CLIO_PREFIX / CLIO_BIN_DIR and the server port with CLIO_PORT. Full install + uninstall reference: install/README.md.

Architecture

For the design rationale (three-tier hierarchy, ARC memory, optimizer layer, IOWarp integration), see docs/CLIO_AGENT_ARCHITECTURE.md.

Documentation

The full, organized index is in docs/README.md. Common starting points:

• Setup & install — release install, source build, smoke test
• Architecture — the three-tier design, ARC memory, and tool gateway
• Contributing — dev environment, quality checks, and where code goes
• Adding tools / experts — extend the agent
• Providers — configure an LM backend
• Roadmap — what's next

Contributing

Start with docs/CONTRIBUTOR_QUICKSTART.md. For guidance specific to AI agents working on the codebase, see AGENTS.md.

License + Citation

BSD-3-Clause.

@software{clioagent2026,
  title  = {CLIO Agent: Autonomous Agent for Scientific Data Management},
  author = {IOWarp Team},
  year   = {2026},
  url    = {https://github.com/iowarp/clio-agent},
}