gauge-sh/gauge-cli

This page introduces an external open-source repository. It is not an HDATF product.

What it does

This developer CLI repeatedly tests whether agents can actually use a product's tools and documentation. It accepts inputs such as Skills, CLIs, SDKs and docs, together with realistic tasks and observable criteria. Outputs retain session transcripts, tool activity, judgments and evidence for improvements. Local mode injects working-tree inputs into fresh Docker sessions and stores evidence and measurements in a private workspace. The user's coding agent writes improvements and analyzes results in local mode. Requirements include Node.js 22.13 or newer, npm, Docker running Linux containers and credentials for the tested agents. The local loop needs no Gauge account, but model usage belongs to the provider account or subscription. Managed execution, automatic analysis and authenticated MCP/API inputs belong to hosted capabilities, with separate identities and storage across modes. Default CLI usage telemetry can be disabled, and web-response fixtures are not complete network isolation, so test data and secret boundaries need review.

License

MIT Permissive. Commercial use and changes are allowed if the copyright notice is kept.

More in this category

• tembo/voxl
  This Rust tool lets a web browser view and control a server or virtual machine desktop. It captures an X11 display on Linux and the running computer's screen on macOS and Windows. The screen is encoded as H.264, sent over a WebSocket and played through the browser's WebCodecs support. It provides mouse and keyboard input as well as clipboard, file transfer and audio features. Agents can use APIs for screenshots, video recording and waiting for screen changes or stillness. The wait implementation tracks changed screen regions and sequence numbers; a still screen does not itself establish task success. Getting started requires a target desktop and platform-specific video and audio libraries, with a protected connection such as HTTPS for remote browser access. It does not create a separate virtual machine, and authenticated viewers can normally control the streamed desktop. The documentation says Wayland and GPU encoding are not yet supported. Its own code is MIT, but the README states that distributed binaries linked with x264, and distributions bundling them, fall under the GPL.
• ivankovic/omnidiff
  OmniDiff compares two files and presents changes according to their content rather than only matching lines. For supported programming languages, it parses syntax trees with tree-sitter and combines subtree matching, residual matching and move recovery to identify insertions, deletions, updates and moves. Images, archives, fonts, translation catalogs and PDFs have content-specific comparisons, while unsupported text uses line comparison and unknown binary content receives a change notice. Results can be viewed in an interactive terminal interface or emitted as plain text or JSON for other tools. Git and Jujutsu integrations require their own configuration, and the JSON format supplies file paths and changed ranges rather than embedding file contents. Linux, macOS and Windows binaries are documented, while building from source needs Rust 1.92 or later and a C compiler. Starting with a pair of local files does not require a model service, and the documented stable integration surfaces are the command line and JSON, not every internal Rust API. Reported performance and accuracy belong to the author's datasets, the stated highlighting accuracy goals are not yet met, and licensing needs review because the README and package declare AGPL with a separate commercial option.
• storytold/craft-anything
  Craft Anything is a Rust application for inspecting websites, applications and games and having a coding agent create a separate project from the findings. Its intended users are people who want to understand software or build their own version, supplying a target address, files, screenshots and a natural-language request. The default Clean room mode uses externally visible material without reading the original code or page source, while Deep mode permits internal inspection of authorized targets. It follows a study, build, compare and refinement workflow, and website comparisons cover multiple screen widths. Each project keeps the request and history, captures, generated application, comparison results, evidence and open questions in separate files or folders. The same engine is available through a desktop interface, command-line tools and MCP, a protocol that lets agents call tools. Building primarily uses the user's signed-in Claude Code, while analysis can use Anthropic, OpenAI, OpenRouter or an OpenAI-compatible local server. The documented source setup requires Rust 1.97 or later and a local coding agent, and provides desktop and command-line entry points. It is early alpha: Codex and API-based builds lack live verification, some inspection tools are partial, game self-tests rely on the generated game's own report, and use must stay within examination and asset rights without bypassing DRM or logins.
• l0ng-ai/openidea
  OpenIDEA is a desktop development environment that brings IntelliJ-style shortcuts, tool windows, and global search to macOS, Windows, and Linux. Developers open source projects to edit code, inspect changes, debug, use Git and terminals, and access database and HTTP clients. Its interface uses React, TypeScript, and CodeMirror inside Tauri, with a Rust backend rather than an entirely Rust-written interface. The agent manager sends prompts and file context to external coding agents through adapters and ACP, an editor-to-agent communication protocol. Tasks can use the project directory or a separate Git worktree, and persist conversations, status, and changed-file information. Users start with a platform build or the documented Node, pnpm, and Rust development setup, while agent executables, service authentication, and some language tools are separate requirements. Conversation rewind does not restore modified files, and the implementation includes modes that can automatically approve tool requests. The project is Apache-2.0 with separate third-party notices, and its author's startup benchmark is not an equivalent-work comparison or an independently verified performance result.
• microsoft/TokenEconomizer
  TokenEconomizer is a plugin intended to reduce waste when coding agents repeatedly work in the same repository. It takes session logs, tool usage and failure history and produces repository-specific learned facts, tool exclusion proposals or settings, and redirects for repeated failures. Micro-skills inject facts confirmed across sessions into later context and drop contradicted facts. Tool pruning collects evidence of nonuse, excludes tools through host settings and allows restoration. Failure-loop detection consults a judge model and suggests another approach, but does not stop work itself. It requires Python 3.10 or newer, an authenticated supported coding agent and plugin installation, with extra model calls handled through that agent's account. Configuration code disables micro-skills by default while enabling tool pruning and failure-loop nudges, so installation does not enable every learning feature. Claude Code, Codex, GitHub Copilot CLI and Pi are supported, opencode is experimental, and tool pruning does not act on Pi or opencode. State remains local, and the authors' cost experiments do not guarantee equivalent savings for individual workloads.

Only repositories in the ranked Trendshift lists are included, and the lists are used only to find candidates. We do not copy their ranks. Descriptions, licenses and star counts come from each GitHub repository. The notes are our own reading. We have not tested these projects, and a place on a trending list does not prove quality.