CLI-Anything: Making ALL Software Agent-Native
CLI-Anything: Making ALL Software Agent-Native
![]()
Today’s software serves humans. Tomorrow’s users will be agents. CLI-Anything is the bridge that turns that future into a practical present: a universal, agent-native pathway to control, automate, and orchestrate any software that ships with a codebase. This is not a collection of fragile scrapes or brittle UI automations. It’s a principled, production-grade approach that treats real applications as first-class backends for AI-driven workflows.

CLI-Anything is built around a single, ambitious idea: any professional tool can be exposed to AI agents through a coherent, production-ready Command Line Interface (CLI). Agents read, reason, and act through structured commands that map directly to real software backends — whether that’s generating a document with LibreOffice, rendering a 3D scene in Blender, or stitching a video with Shotcut. The result is a suite of ready-to-use harnesses, each generated by an automated pipeline that preserves full capability, accuracy, and determinism.
What is CLI-Anything?
CLI-Anything is a system and methodology for turning software with a codebase into agent-native CLIs. It accomplishes this with a seven-phase pipeline, a unified REPL interface, and a firm commitment to authentic software integration. The project centers on the idea that agents should interact with real backends, not “toy” implementations or simplified APIs.
Key ideas at a glance:
- Structured, composable commands that mirror human-invoked workflows, but in machine-friendly JSON and human-readable forms.
- A production-grade pipeline that analyzes, designs, implements, tests, documents, and publishes CLI harnesses for real software.
- An agent-first design that outputs deterministic results and clean JSON for downstream reasoning, with friendly human-readable modes for debugging.
- A hub of community-built CLIs, powered by a meta-skill that lets agents discover, install, and run the right CLI for a given task.
The project embraces a broad ecosystem of platforms and agents, from Claude Code and Pi Coding Agent to OpenClaw, Codex, GitHub Copilot CLI, and beyond. The CLI-Hub registry makes it possible for agents to browse, install, and execute the exact CLI they need, with live updates reflecting changes in the catalog.
To illustrate the breadth, the project ships with a growing set of ready-to-use CLIs, including cli-anything-gimp, cli-anything-blender, cli-anything-drawio, cli-anything-zotero, cli-anything-obs-studio, cli-anything-kdenlive, cli-anything-shotcut, cli-anything-openscreen, cli-anything-zoom, cli-anything-musescore, cli-anything-drawio, and many more. Each harness results from a disciplined, repeatable process designed to cover production-quality usage, not just a demo.
Core Principles and Goals
- Authentic Software Integration: The CLIs call the real software backends (e.g., Blender, LibreOffice, GIMP) to render results with no compromises.
- Agent-Native Design: Every command emits structured JSON output alongside human-friendly formatting, enabling seamless agent consumption.
- Fully Automated Pipeline: The seven phases run automatically, producing SKILL.md files and documentation that reflect the actual capabilities of each harness.
- Zero-Config Installation: A simple pip install -e . puts cli-anything- on PATH; agents discover tools via standard which commands.
- Robust Testing: The project ships with thousands of tests across dozens of backends to ensure reliability in production-like conditions.
Quick Start
Prerequisites
- Python 3.10 or newer
- Target software installed (GIMP, Blender, LibreOffice, or your own application)
- An AI agent that can drive CLIs (Claude Code, Pi Coding Agent, OpenClaw, OpenCode, Codex, GitHub Copilot CLI, etc.)
Pick Your Platform (examples)
- Claude Code: Install the CLI-Anything marketplace plugin, then generate a CLI for a target app with a single command.
- Pi Coding Agent: Install the Pi extension globally so /cli-anything works across Pi projects; generate CLIs from any repository.
- OpenCode, OpenClaw, Codex, Copilot CLI, and others can participate via their respective integration points.
A representative workflow
- Generate a CLI harness: /cli-anything ./gimp
- Work through the 7 phases automatically (Analyze → Design → Implement → Plan Tests → Write Tests → Document → Publish)
- Use the REPL or the generated CLI in JSON mode for agent consumption
Images from the Input show the typing and teaser visuals that accompany the CLI-Anything experience, reinforcing how quickly a guided agent can build an autonomous CLI harness.
The 7-Phase Pipeline: From Codebase to Production CLI
CLI-Anything uses a deterministic, repeatable workflow to transform a software project into a production-ready CLI harness. The seven phases ensure end-to-end coverage, from understanding the codebase to delivering a usable CLI and a published artifact.
- Analyze: The agent scans source code, maps GUI actions to APIs, and identifies the capabilities that must be exposed to the CLI.
- Design: The harness architect designs command groups and a state model, deciding how output formats will look and how interactions will flow.
- Implement: The CLI is built using Click, with a REPL (ReplSkin) and JSON output. The system ensures undo/redo paths and a clean, testable interface.
- Plan Tests: A TEST.md is drafted to capture unit and end-to-end test plans aligned with actual usage.
- Write Tests: A comprehensive test suite is implemented to validate core functionality and end-to-end workflows.
- Document: Documentation pages, examples, and usage notes are produced to aid both humans and agents.
- Publish: The harness is packaged, installed to PATH, and made discoverable by the registry so agents can find and use it autonomously.
The pipeline is designed to be repeatable and non-destructive. If a subsequent refinement is needed, you can run refine commands to expand coverage, focusing on specific functionalities (for example, “batch processing and filters” in image workflows).
The system also ships a canonical SKILL.md for each harness, generated from the CLI’s metadata. This skill description supports agent discovery and orchestration, ensuring tools are both machine-readable and approachable for human operators.
Architecture and Design: How CLI-Anything Keeps It coherent
A single, cohesive architecture ties together all CLIs, across backends and platforms.
- Fully Automated 7-Phase Pipeline: Architecture supports end-to-end harness generation, with test plans and SKILL.md generation baked in.
- Authentic Software Integration: Each CLI calls the actual software backends to render results (ODF for LibreOffice, MLT XML for video, etc.). No mock backends; no compromises.
- Dual Interaction Model: Every CLI supports a stateful REPL for interactive agent sessions and a standard subcommand interface for batch scripting and pipelines.
- Unified REPL Experience: A common REPL skin (ReplSkin) provides consistent banners, prompts, command history, and progress indicators across all CLIs.
- Agent-Native Outputs: Each command includes a --json option, delivering machine-friendly results suitable for agents, complemented by human-readable tables for interactive use.
- Zero-Config Installation: pip install -e . enables immediate PATH access; agents discover tools using standard commands like which or path checks.
- Zero-Dependency Philosophy: The system uses real software as the backend. Missing dependencies trigger test failures rather than silent degradation, preserving authenticity.
A clear architectural image in the input highlights the relationships and modular boundaries, showing how harnesses reside under cli_anything.* namespaces and how the SKILL.md files anchor agent discovery.
Project Structure: Where Everything Lives
cli-anything/ is organized into a clean, scalable monorepo. At the top level you’ll find:
- assets/ for images and media (icon, teaser, demos)
- The plugin and harness directories, including cli-anything-plugin and many agent-harnesses (gimp, blender, inkscape, audacity, libreoffice, zotero, mubu, obs-studio, kdenlive, shotcut, openscreen, zoom, musescore, drawio, eth2-quickstart, mermaid, anygen, comfyui, notebooklm, dify-workflow, adguardhome, ollama, unimol-tools, videocaptioner, godot, s&box, lldb, nsight-graphics, unrealinsights, cloudanalyzer, qgis, and more)
- HARNESS.md and SKILL.md templates
- A collection of CLI harnesss under cli_anything.* namespaces for separation and conflict-free packaging
- A compact, readable READMEs and PUBLISHING guides to help new contributors
The architecture image in the content gives a snapshot of how all pieces fit together, from pipeline stages to SKILL generation and the registry-driven discovery.
Real-World Demos: Agents Building Real Art, in Real Time
The CLI-Anything project isn’t just theory; it shows live, artifact-producing demos across multiple domains. Here are a few highlights, with visuals from the input to anchor the narrative:
FreeCAD — Curiosity Rover via Preview, Live Preview, and Trajectory
Harness: cli-anything-freecad
Preview Stack: preview, preview live, trajectory.json
Artifact: An agent-built Curiosity-like rover, with live previews and a trajectory history that documents how the rover evolved
Visual: FreeCAD Curiosity rover demo driven by preview, trajectory history (gif)
Blender — Orbital Relay Drone via Preview, Live Preview, and Trajectory
Harness: cli-anything-blender
Preview Stack: preview, preview live, trajectory.json
Artifact: An agent-built orbital relay drone, grown stage-by-stage with a live head-tracking session
Visual: Blender orbital relay drone preview-trajectory (gif)
Draw.io — HTTPS Handshake Diagram
Harness: cli-anything-drawio
Artifact: A complete HTTPS handshake diagram (.drawio and .png)
Visual: Draw.io CLI demo GIF, followed by the final handshake diagram
Slay the Spire II — Game Automation
Harness: cli-anything-slay-the-spire-ii
Artifact: An automated gameplay session
Visual: Slay the Spire II CLI gameplay GIF
VideoCaptioner — Auto-Generated Subtitles
Harness: cli-anything-videocaptioner
Artifact: Captioned video frames with bilingual rendering
Visuals: Before/after frames showing captioning results
Each demo demonstrates a consistent pattern: the agent drives a complete, end-to-end pipeline via the CLI harness, generating artifacts that are directly usable in the target domain. The visuals in the input—a sequence of demos, GIFs, and stills—show the live progression from initial commands to final outputs. They are not mere proofs of concept; they depict production-grade pipelines with replayable histories and verifiable renders.
The CLI-Hub and the Meta-Skill: Autonomy at Scale
CLI-Hub is a central registry designed to empower agents to discover and install CLIs with zero human intervention. The concept is simple but powerful: a meta-skill lets agents browse the catalog, install the best-fitting CLI, and execute tasks end-to-end.
- Install in one command (example): openclaw skills install cli-anything-hub
- The agent then searches the catalog, installs the appropriate CLI, and uses it to complete the task
- The catalog auto-updates when registry.json changes, ensuring new community CLIs appear automatically
For Claude Code users, the same discovery capability can be achieved by embedding the SKILL.md into a project or skills folder, enabling automatic agent discovery and orchestration.
This hub-and-skill approach transforms the way agents interact with software: they no longer rely on handcrafted wrappers or brittle automation. They leverage a living registry of real CLIs, all of which are generated through the seven-phase process and validated with thousands of unit and end-to-end tests.
The Vision: Building Agent-Native Software
CLI-Anything’s long-term aim is straightforward and transformative:
- Universal Access: Every software becomes instantly agent-controllable via a well-defined CLI.
- Seamless Integration: Agents control applications without bespoke APIs, UIs, or wrapper code.
- Future-Ready Ecosystem: A scalable path to transform human-designed tools into agent-native assets with a single command.
The project’s roadmap, test history, and growing catalog demonstrate a clear, practical path toward that future. The manifest of features, demonstrations, and community involvement makes a compelling case for why CLI-Anything is more than a provocative concept—it’s a usable, extensible platform.
What Can You Do with CLI-Anything?
Let Agents Take Your Workflows
Drop a codebase into /cli-anything to generate a full harness for tools like GIMP, Blender, Shotcut, LibreOffice, OBS Studio, and more.
Instantly obtain a production-grade CLI that agents can drive from start to finish.
Unify Scattered APIs into One CLI
Feed documentation or SDKs into /cli-anything to wrap disparate endpoints into a coherent, stateful CLI with consistent commands and outputs.
Replace or Supercharge GUI Agents
A generated CLI can replace brittle GUI automation and empower fully automated task synthesis, evaluation, and benchmarking via code.
Bullet points above echo the project’s emphasis on pragmatic, scalable agent-controlled workflows that preserve full capabilities of the underlying software.
Images from the Input provide a visual spine for this narrative: the typing demo, teaser visuals, and the live-demo GIFs illustrate how quickly a real, fully capable CLI harness can be produced and reused by AI agents.
Test Results and Quality Assurance
CLI-Anything has undergone rigorous multi-layered testing across a broad spectrum of applications. The project reports:
- 2,280 tests passing in total (unit + end-to-end, with Node.js tests included)
- A 100% pass rate across all harnesses
- Demonstrable coverage across 18 major applications, including creative, productivity, diagramming, and development domains
This testing backbone is essential for agents to rely on these CLIs in real tasks where reliability matters. It’s not just a showcase; it’s a production-grade quality assurance story that underpins trust in agent-driven automation.
Getting Involved: Contribute and Extend
CLI-Anything is designed to be extensible and community-driven:
- New targets: Generate CLIs for any software with a codebase and contribute harnesses via the publishing guidelines.
- Methodology improvements: Propose changes to HARNESS.md or the seven-phase methodology to encode new lessons.
- Plugin enhancements: Add commands, improve validation, or expand test coverage with new E2E scenarios.
- Test coverage: Expand scenarios, edge cases, and workflow tests to broaden reality-grounded validation.
The project’s governance encourages open participation and contributions that accelerate the agent-native future.
Roadmap, Limitations, and Future Work
- Roadmap highlights: support for more tool categories (CAD, DAW, IDE, EDA, scientific tools); a benchmark suite for agent task completion; community-driven harnesses; deeper integrations with agent frameworks beyond Claude Code.
- SKILL.md generation: every generated CLI ships with a canonical SKILL.md, enabling easier discovery and orchestration by agents.
- Limitations: the system relies on strong foundation models; some closed-source or binary-only targets may require additional adaptation; iterative refinement is often needed to achieve production-level coverage.
An honest view of limitations is paired with a practical path forward: more platforms, broader category coverage, and deeper agent integration.
A Final Note: The Future Is Agent-Native
CLI-Anything represents a practical, scalable approach to the age of AI agents. It claims no magic, only a disciplined pipeline that consistently turns codebases into agent-ready CLIs, with authentic backends, deterministic results, and robust test coverage. It emphasizes a universal interface that AI agents can understand and reason about, while preserving the richness and power of the original software.
If you’re curious about trying it, you can explore the CLI-Hub, generate harnesses for your software, and perhaps contribute a new CLI harness to the growing catalog. The visuals and demos in this post reflect ongoing progress across multiple domains—proof that the agent-native future is not a distant dream, but a tangible, working ecosystem you can participate in today.
Visuals and Demos Recap
- CLI-Anything typing demo: a quick look at how agents type commands and manage sessions
- Visual: asset/cli-typing.gif
- Teaser image for the project
- Visual: asset/teaser.png
- Real-world demos and artifacts
- FreeCAD Curiosity rover: asset/demos/freecad-curiosity-preview-trajectory.gif
- Blender orbital relay drone: asset/demos/blender-orbital-relay-drone-preview-trajectory.gif
- Draw.io HTTPS handshake diagram: asset/demos/drawio-demo.gif; final PNG: asset/demos/drawio-https-handshake.png
- Slay the Spire II gameplay: asset/demos/slay-the-spire-ii-gameplay.gif
- VideoCaptioner before/after frames: asset/demos/videocaptioner-before.png and asset/demos/videocaptioner-after.png
- CLI-Anything Architecture
- Visual: asset/architecture.png
These visuals anchor the narrative and demonstrate, visually and concretely, how AI agents can generate, run, and verify complete CLIs that control real-world software.
If CLI-Anything resonates with your goals, imagine a world where every tool you rely on can be orchestrated by an agent through a stable, predictable CLI. The trajectory is clear: a future where agent-native software is no longer a novelty but a standard, enabling faster, more reliable automation across creative, technical, and enterprise domains. Welcome to that future. Welcome to CLI-Anything.
Enjoying this project?
Discover more amazing open-source projects on TechLogHub. We curate the best developer tools and projects.
Repository:https://github.com/HKUDS/CLI-Anything
GitHub - HKUDS/CLI-Anything: CLI-Anything: Making ALL Software Agent-Native
Today’s software serves humans. Tomorrow’s users will be agents. CLI-Anything is the bridge that turns that future into a practical present: a universal, agent-...
github - hkuds/cli-anything