Agent frameworks#Device control

CLI-Anything: generate CLIs so agents can drive real software

Generates structured CLIs for GIMP, Blender and LibreOffice so agents drive real software backends; Apache-2.0, 51k stars, CLI-Hub one-command installs.

Project facts

GitHub Ecosystem
Repositorygithub.com/HKUDS/CLI-Anything
License
Apache-2.0
Language
Python
Stars
51,411
Data checked
2026-10-03

Snapshot figures reflect the check date and may change over time.

AI agents write decent code but remain clumsy at using software: asked to drive Blender, GIMP or LibreOffice, they either bypass the GUI with a hobbled API or fall back to vision-model coordinate clicking — brittle, slow and hard to reproduce. CLI-Anything from HKUDS — the team behind LightRAG — takes a third path: generate a command-line interface an agent can understand, compose and verify, while the actual rendering and file generation stay with the real software backend instead of a simplified stand-in. Created March 8, 2026, the project counted 51,411 stars as of October 3, written in Python under Apache-2.0. The WeChat account 架构师修行之路 covered it on October 3.

CLI-Anything demo: an agent drives real FreeCAD through the generated CLI to build a Curiosity rover model, with preview, live view and trajectory history side by side

Core features

  • Seven-phase HARNESS generator: study the software’s code and data model, design the CLI, implement it, write the test plan and scripts, complete docs, package and publish. Output lands in agent-harness/ — a Click CLI, core module, interactive REPL, JSON output, tests and install docs. For agents, --help is the self-describing entry point and --json the stable output channel.
  • Authentic software integration: the CLI only converts operations into structured commands; rendering and file generation remain with the real applications. The README lists this as a named principle — generated project files are genuine formats, not simulated effects.
  • CLI-Hub ecosystem: pip install cli-anything-hub browses, searches, installs and upgrades community-built harnesses from one command. The registry goes well beyond the usual suspects — alongside GIMP, Blender, Inkscape, LibreOffice and Kdenlive there are FreeCAD (258 commands), Zoom, Obsidian, Zotero, QGIS, Godot, n8n and Safari, plus a Slay the Spire II gameplay-automation harness.
  • refine loop: /cli-anything:refine <path> [focus] runs gap analysis between the software’s full capabilities and the current CLI, then adds commands, tests and docs incrementally and non-destructively.
  • SKILL.md generation: every harness ships a skill file, and a CLI-Hub meta-skill (one npx skills add) lets agents discover and install the CLI a task needs on their own.

Typical use cases

  • Batch content production: agents drive Inkscape for graphics, Kdenlive for video and LibreOffice for documents, producing real-format files that flow into downstream pipelines.
  • Engineering and research workflows: batch jobs in FreeCAD modeling or QGIS cartography, where the agent writes scripts and the real software produces results.
  • Testing and automation: give API-less desktop software an assertable command interface, with test cases generated alongside the harness.

Quick start

For ready-made tools (the target software must be installed):

pip install cli-anything-hub
cli-hub search image
cli-hub install gimp
cli-hub launch gimp

For software the registry doesn’t cover, generate a harness — with Claude Code:

/plugin marketplace add HKUDS/CLI-Anything
/plugin install cli-anything
/cli-anything ./gimp

Python 3.10+ required; iterate with /cli-anything:refine ./gimp to grow coverage.

Summary

This is for people building automations around real files and real projects; anyone expecting one-click full automation will be disappointed — both the README and the coverage are explicit that the target software still has to be installed, missing backends can’t be conjured into existence, and a generated harness is not complete coverage until tests and refine passes say so. Apache-2.0, friendly to commercial use. For a different route to the same goal, see our earlier entry on phone-harness: screen OCR plus coordinate taps for phones, where no CLI can be generated — the two approaches complement each other, so pick by target software.