CLI-Anything's bet: every GUI app gets a terminal, and agents get the keys

The pitch on the CLI-Anything GitHub repository is blunt: “Today’s Software Serves Humans. Tomorrow’s Users will be Agents.” The project, from the Hong Kong University of Data Science (HKUDS), generates command-line interfaces for desktop applications that otherwise have no terminal story, then packages those CLIs with AI-discoverable skill definitions. The result is a registry of agent-native harnesses covering Blender, FreeCAD, Zoom, Obsidian, QGIS, Godot, and dozens more. It is one of the most direct attempts yet to answer a question every AI builder is circling: if agents are the next users, how does software get built to serve them?

The project’s own numbers tell the story of how fast this is moving. Since the CLI-Hub package manager launched on April 10, 2026, the team has merged harnesses at a furious clip. The news log shows FreeCAD landing with 258 commands across 17 groups on March 24, MuseScore on March 22, Zotero on March 26, and a wave of merges in mid-April that included QGIS, UniMol Tools for molecular modeling, and an Unreal Insights CLI with background capture session control. The Joplin harness shipped with a 134-test validation baseline. The Obsidian CLI carries 48 unit tests and 7 end-to-end tests. This is not a toy script collection; it is a quality-gated integration effort.

The technical mechanism matters. Each harness is generated through a seven-phase process that produces a HARNESS.md file, a README, and a SKILL.md that agents can discover autonomously. Since April 18, all skill files live under a top-level skills/ directory, installable with a single npx skills add HKUDS/CLI-Anything --skill <skill-name> command. The CLI-Hub meta-skill, launched March 23, lets an agent browse and install CLIs on its own. That is the quiet revolution here: the agent does not need a human to set up the tooling. The agent can find the CLI, install it, read the skill file, and start driving the application.

The project’s “Why CLI” section makes the argument concisely. Text commands match LLM format and chain for complex workflows. --help flags provide self-describing documentation agents can discover. Structured JSON output eliminates parsing complexity. Claude Code runs thousands of real workflows through CLI daily, the README notes. These are the properties that make terminals the natural substrate for agent control, and CLI-Anything’s bet is that they can be retrofitted onto any application, no matter how GUI-centric its design.

The security posture deserves attention. The project has been hardening aggressively. A GIMP Script-Fu path injection vulnerability was fixed on March 14. Zoom token permissions were tightened on March 25. The Sketch CLI token-file handling was hardened against path traversal and symlink escapes on May 21. A May 19 pass routed XML, SVG, ODF, MLT, MusicXML, and CSL parsing through defusedxml to block untrusted input. DomShell, the browser CLI, got URL validation and DOM sanitization on April 7. This is the unglamorous work that makes agent-native software safe to run at all, and it is exactly the work that gets skipped when a demo video is the deliverable.

What is genuinely new here is the scope of the ambition. Other projects have wrapped individual apps in agent-friendly interfaces. CLI-Anything is attempting a universal adapter layer, a protocol for making any software agent-native. The registry already spans video editing (Kdenlive, Shotcut, Openscreen with 101 tests), music production (Rekordbox with SQLCipher write guards, MuseScore), 3D printing (3MF mesh inspection and repair), and even a Slay the Spire II harness for the deck-building roguelike. The inclusion of a game is a signal: the project is not just about productivity software. It is about total coverage.

The economics here are worth dwelling on. Every harness in the registry represents a small integration project that a company like Anthropic or OpenAI would otherwise have to build, maintain, and secure in-house. By open-sourcing the generator and the registry, HKUDS is effectively crowd-sourcing the agent-native retrofit of the world’s software. The community contribution model is explicit: build a harness, get it reviewed and merged, gain access as a community contributor. The wishlist feature lets users request specific software. This is a classic open-source flywheel, and it is spinning fast.

There are real questions the project does not answer. The quality bar varies by harness, even with the test baselines. A 134-test Joplin harness is not the same as a 41-test Calibre harness, and neither guarantees compatibility with every future version of the underlying application. Desktop apps change their internals, and a harness that breaks on update is a support burden. The project’s reliance on --help output for self-description is clever but fragile; not every application exposes complete documentation through its CLI, and some of the generated harnesses have had to add explicit constraints to their skill files to keep agents from doing destructive things.

The Rekordbox harness is a useful case study in how the project handles risk. It ships with guarded SQLCipher write paths, backup-required forced writes, and smoke coverage. The Zoom harness was hardened so downloads come from direct URLs rather than whatever link an agent might find. These are the details that separate a demo from a tool that can be trusted with real user data. The project clearly understands this, and the security fixes in the news log suggest a team that treats the threat model seriously.

For AI builders, the implication is direct. The agent-native software stack is being assembled right now, and much of it is being built by volunteers in public repositories rather than by the frontier labs. CLI-Anything is one of the most active projects in this space, and its registry is a map of what agents will be able to control in the near term. The labs are building the models and the agent frameworks; the open-source community is building the interface layer that lets those agents actually do things in the world.

The project’s own trajectory suggests where this is heading. The CLI-Hub v0.2.0 release on April 15 added support for public CLIs from multiple install sources, including pip, npm, brew, and bundled system tools, backed by a public registry file. The May 27 update handled Python package-date detection for registry automation. Every release makes the registry more automated, more self-maintaining, and more capable of running without human intervention. The meta-skill that lets agents discover and install CLIs autonomously is the endgame: a software ecosystem that maintains its own agent-facing interfaces.

The open question is whether the model scales to the long tail of software. There are thousands of desktop applications, and each harness requires maintenance. The community contribution model helps, but it depends on sustained interest. The project’s momentum through May 2026 is real, with merges landing nearly every day. Whether that pace holds through the end of the year will determine whether CLI-Anything becomes the universal adapter layer it aspires to be, or a well-executed proof of concept for a smaller set of tools.

The trendshift badge on the README and the daily merge cadence suggest the former is more likely. The agent-native software stack is being built in public, one harness at a time, and HKUDS has positioned itself at the center of that effort. The terminal is having its moment again, not as a relic of computing’s past, but as the interface layer for computing’s next users.