CLI-Anything: Make Any Software AI Agent-Ready

Background and Context

Most desktop software was built for people. Buttons, menus, drag handles and dialog boxes all assume an operator with eyes and a mouse. AI agents break that assumption. When a model drives an application through screenshots and simulated clicks, the process is slow and brittle, and a small interface redesign can break it. CLI-Anything, an open-source project from the Data Intelligence Lab at the University of Hong Kong (HKUDS), targets this gap. Its stated goal is to make almost any software usable by agents without a graphical interface in the loop.

The idea is simple. Agents are good at reading text, writing commands and parsing structured output. So the project gives each application a command-line layer. Work that once needed a mouse becomes a command that can be scripted and combined with other commands. The project sits between classic software automation and newer agent tool protocols. It aims at legacy desktop programs that have no open API and no ready-made MCP server.

Interest is visible. The repository has collected more than 54,000 stars on GitHub, which suggests many developers feel the same problem.

Deep Analysis

What does CLI-Anything actually ship? It ships two things. The first is a method for generating and maintaining CLI wrappers for different programs. The second is CLI-Hub, a registry and installer for those wrappers. A developer runs pip install cli-anything-hub, then uses cli-hub install to fetch a specific tool. The wrappers work with agent environments including Pi, OpenClaw, nanobot, Cursor and Claude Code. A single CLI is therefore not tied to one model or one product. The project activity shows three design priorities. The first is security. The Rekordbox CLI protects its SQLCipher write path and forces a backup before any write. The Sketch CLI fixed a path traversal flaw and a symlink escape in its token file handling. The second is engineering discipline. The Joplin CLI was submitted with 134 tests as a verification baseline. The Calibre CLI carries 41 unit tests plus end-to-end evidence against a real Calibre install. The third is breadth. Notes, e-books, music management, 3D meshes, geographic information systems and the Unreal Engine editor all appear in the registry or in proposals.

Compared with RPA tools or screenshot-based computer control, the approach swaps one kind of uncertainty for another. Guessing what a screen shows becomes calling a named command. Failures are easier to locate, because a command either succeeds or returns an error the agent can read. How does a team start using CLI-Anything? First check CLI-Hub for an existing wrapper for the target application and install it. If none exists, the project offers two routes. A developer can submit a wish-list request for a program or service. Or the developer can sign up as a contributor, build a new CLI harness, and have it reviewed and merged into the public registry. The demos show agents using generated CLIs with previews, live previews and trajectory loops. The outputs include CAD models, 3D scenes, charts, game play and subtitles. The point is that the agent sees intermediate results and corrects its next step. Documentation exists in English, Chinese, Japanese and German.

Industry Impact

CLI-Anything tries to turn software agent-readiness from scattered vendor effort into shared, crowd-built infrastructure. If the registry keeps growing, agent developers need not rebuild integrations for every application. Engineering teams could fold internal desktop tools into agent workflows faster.

The pace of contribution is already visible. In late May 2026 alone, the project merged or proposed a Hermes orchestration skill and wrappers for ArcGIS Pro, Obsidian Agent CLI, Joplin, Rekordbox, Calibre, 3MF, MiniMax and UEAtelier. Existing wrappers for Zoom, Obsidian and LibreOffice also received quality and security fixes.

The risks are real. Wrapper quality will vary. Community CLIs differ in stability, permission boundaries and version support, and users must judge them. Tools that write to databases or user files deserve a backup first. Letting an agent run commands also widens the attack surface. The path traversal and token issues already seen in the Sketch CLI show that review must be continuous. Finally, an application update can break a wrapper, and the maintenance cost falls on the community.

Outlook

Three questions will shape the project. First, can the review and signing process for registry entries keep pace with growth? A hub that installs commands for agents is only as trustworthy as its vetting. Second, will preview and feedback loops for graphical programs mature beyond the current demos? CAD, 3D and music software need the agent to see results, not only send commands. Third, how will CLI-Anything divide work with other agent integration standards such as MCP?

For developers and engineering teams who want agents to drive CAD, 3D, document or music software, the practical step is small. Check CLI-Hub for the program you use, test the wrapper on disposable data, and back up anything it can write to. If the registry handles security and maintenance well, it could become a common entry point for agents to operate real-world software.