OmniGet Desktop Agents & Downloader
Run coding agents with approval and undo, then download, transform, and manage media locally.
Per-dimension scores and reasoning
The README describes workspace path confinement, pre-write command or diff approval, per-segment validation of persistent rules, cancellation of pending approvals, one-call external-directory grants, and turn-level shadow-Git checkpoints and undo. This provides strong evidence for confirmation, external-effect visibility, and recovery. A key-management test verifies preservation of existing variables and mode 0600, while network use, remote pruning, and the room relay receive some disclosure. Deductions apply because shell sandboxing is only specified for macOS, while arbitrary shell commands, webhooks, and remote services remain available; privacy and token handling rely mostly on README assertions plus one narrow test; broad dependency ranges are present without demonstrated vulnerability scanning, auditing, or supply-chain policy; and authorship, third-party attribution, and maintainer identity are only lightly evidenced.
The README, package metadata, and CI are broadly consistent about version, supported platforms, test commands, and product scope. CI covers Rust tests on three platforms, frontend checks, unit tests, locale consistency, Debian compatibility, and Windows/Linux launch smoke tests; focused tests cover key updates and downloader retry failures. Deductions apply because many agent, queue, recovery, ACP, MCP, and performance claims lack corresponding implementation files or targeted tests in the supplied evidence, macOS has no launch smoke job, and failure recovery for automatic installation and updates is not shown thoroughly.
The material clearly distinguishes downloader, coding-agent, local-model, CLI/ACP, job, loop, schedule, webhook, MCP, and visitor scenarios, with cross-platform and local-first paths. Boundaries include a single attached workspace, budgets, concurrency, permission levels, loop stop conditions, visitor restrictions, and beta labels. Deductions apply because trigger behavior is principally documented rather than supported by parser or routing code, permission and sandbox guarantees differ across operating systems, and compatibility limits for remote providers, external CLIs, and system tools are incomplete.
The README has a strong contents structure, task-oriented tables, stable LLM/Jobs/Loops/World terminology, command examples, and entry points for installation, FAQ, building, and contribution. Package version 0.10.0, a version script, and release-driven Homebrew automation establish an update path. GPL-3.0 is consistent across metadata, badge, and the full LICENSE, justifying full license credit. Deductions apply because the supplied excerpt does not contain the actual installation, FAQ, build, or changelog text; limitations are scattered rather than comprehensive; and the unverified publisher is not accompanied by clear maintainer ownership, support commitments, or a security contact.
The GUI, diff review, one-click undo, durable jobs, loops, logs, cost reporting, local models, and cross-platform packaging describe a usable end-to-end workflow. Combining download, media, and agent-control capabilities offers substantial marginal value. Deductions apply because the static evidence does not show representative completed outputs, full interaction results, or user acceptance evidence, while 156 tools and broad integrations imply installation, attack-surface, and cognitive costs; performance and time-saving figures remain author-reported.
Many claims include concrete mechanisms, state names, tool names, commands, versions, and CI steps. Package metadata and workflows corroborate versioning, platform, and test-entry claims, while focused tests support key-file permissions and retry behavior. Deductions apply because central safety, privacy, undo, crash-recovery, performance, tool-count, and protocol-compatibility claims mostly appear only in the README without matching source or test evidence here, and promotional claims, author measurements, and demonstrated facts are not consistently separated.
- This is a low-confidence static review; the application, tests, agents, downloader, and updater were not executed.
- Verify sandbox and authorization boundaries independently before enabling shell access, persistent allow rules, webhooks, external MCP servers, or remote models, especially outside macOS.
- Key claims such as no telemetry, local file retention, whole-turn undo, and automatic crash recovery require source audit or reproducible testing.
- The dependency surface is large and many versions use broad ranges; inspect lockfiles, Rust dependencies, downloaded binaries, update signing, and vulnerability-scan results before deployment.
- Course and media downloading may be constrained by platform terms, copyright, and technological-protection rules; users should confirm authorization.
What does this agent do, and when should you use it?
OmniGet is a GPL-3.0 desktop application for Windows, macOS, and Linux that combines an AI coding-agent harness, a download manager, and a local media toolbox. It places Claude Code, Codex, Gemini CLI, ACP-compatible agents, and local Ollama, LM Studio, or llama-server models in one roster, with permission prompts, diff review, sandboxing, and turn-level undo around file changes and commands. Work can run through a durable SQLite job queue or a Loop that repeats agent rounds until a specified test command passes or configured limits are reached. Its bidirectional MCP support lets agents consume external MCP servers while OmniGet exposes 49 of its own tools to clients including Claude Code, Cursor, VS Code, Goose, and Codex. Beyond software development, it provides a yt-dlp GUI, purchased-course downloaders, and 156 local tools for media, subtitles, PDFs, images, and other workflows, with resulting files kept on the user's computer.
A user selects an agent and attaches a workspace; the built-in harness can then invoke fs_read, fs_list, fs_glob, fs_grep, fs_edit, fs_write, fs_apply_patch, shell_exec, todo_write, kb_search, and kb_write. Before a write, OmniGet presents the command or diff and asks for Allow, Deny, or Always; before the first write in a turn, it creates a shadow-Git checkpoint so Undo can restore the entire turn without touching the project's repository. Jobs persist prompts, logs, state, model, token usage, and reported cost in SQLite, while Loops run a user-supplied check such as npm test or cargo test after each round. agent_delegate sends a bounded task to another roster member and returns its answer, while shared project memory comes from AGENTS.md or CLAUDE.md plus Markdown files under .omniget/kb/. On the download side, OmniGet accepts URLs, magnet links, and torrent files, uses native extractors or yt-dlp, and can pass local results through documented components such as FFmpeg, whisper.cpp, aria2, and gallery-dl; the built-in players and readers consume the generated local files.
- A developer who wants to use Claude Code, Codex, or Gemini CLI without managing terminal sessions can review diffs, approve writes, and undo an entire agent turn in a desktop UI.
- A team leaving a repair task unattended can create a durable Job or run a Loop that keeps modifying the project and executing its test command until the check passes.
- A privacy-conscious individual can connect Ollama, LM Studio, or llama-server and run local agents with the same filesystem, shell, permission, and undo controls.
- A user with several model and CLI subscriptions can have a local coordinator call agent_delegate and send a difficult subtask to another roster member such as Claude Code.
- A Claude Code, Cursor, VS Code, Goose, or Codex user can enable OmniGet's MCP server to access its download, PDF, transcription, image, and workspace-scoped coding operations.
- Someone preserving purchased courses or public research media can download supported Udemy, Hotmart, Kiwify, Rocketseat, or yt-dlp content and then play, transcribe, convert, or organize it locally.
What are this agent's strengths and limitations?
- One roster supports Claude Code, Codex, Gemini CLI, arbitrary ACP commands, local runtimes, and API-backed models, with agent_delegate for cross-agent handoffs rather than a single-provider design.
- Writes are guarded by visible diffs, one-time or stored permission rules, workspace path boundaries, and whole-turn Undo; on macOS, shell_exec explicitly uses seatbelt to disable network access and confine writes.
- SQLite-backed Jobs, test-driven Loops, cron expressions, and local webhooks support durable, scheduled, and externally triggered agent work.
- A built-in MCP client and a 49-tool MCP server allow the same application both to consume external tools and provide capabilities to other development clients.
- The coding harness sits alongside yt-dlp downloads, course handling, transcription, PDF processing, and media conversion, which is useful for workflows spanning acquisition and local processing.
- The product is unusually broad—agents, download management, players, and 156 tools—so teams seeking only a small coding client inherit a larger installation and interface surface.
- Remote models, Claude Code, Codex, private platforms, and many download operations still depend on network access, third-party accounts, API keys, quotas, or changing site behavior; only the local-model path is described as offline.
- Sandbox parity is not established across operating systems: the source specifically documents macOS seatbelt but does not describe equivalent Windows and Linux shell isolation in the supplied material.
- The builds lack a paid code-signing certificate, so Windows SmartScreen and macOS Gatekeeper can interrupt first launch; macOS may require manual xattr and codesign commands.
- Linux AppImage users may need FUSE 2 and extra GStreamer packages; absent media plugins can produce silent or black playback or terminate WebKitGTK's web process.
How do you install or deploy this agent?
On Windows 10 or 11, download the x64 installer, portable executable, or MSI from Releases, or run winget install -e --id tonhowtf.OmniGet. On macOS 10.15+, download the Apple Silicon or Intel DMG, or run brew install --cask tonhowtf/tap/omniget; because the app lacks a paid signing certificate, first launch may require xattr -cr /Applications/omniget.app followed by codesign --force --deep --sign - /Applications/omniget.app. Linux builds are provided as amd64/arm64 DEB, RPM, and AppImage packages; newer Debian or Ubuntu installations may need sudo apt install libfuse2 for the AppImage. At first launch, select a language and theme, then let OmniGet install and verify yt-dlp and FFmpeg. For AI work, configure at least one execution source: an authenticated Claude Code, Codex, or compatible ACP CLI; a local Ollama, LM Studio, or llama-server instance; or an OpenAI, Anthropic, or OpenAI-compatible API key in settings.
How do you use this agent?
In the desktop UI, open LLM → Accounts to add Claude Code, Codex, Gemini CLI, or another ACP command; connect local runtimes under LLM → Models. Open LLM → Chat, select an agent, attach a workspace, and describe the requested change; inspect each command or diff and choose Allow, Always, or Deny, using Undo if the turn should be reverted. Create persistent work under LLM → Jobs; for a Loop, also provide the check command and round or time limit. Against a running app, the documented CLI examples are omniget agent run "Fix the failing test in src/cart.js" --agent claude-code --workspace . and omniget agent loop "Make the tests pass" --agent omni --workspace . --check "npm test" --rounds 5. To serve tools to an external client, enable Tools → AI → MCP server and paste the generated configuration snippet for Claude Code, Cursor, VS Code, Goose, or Codex. For downloads, paste a URL, magnet link, or torrent on the home screen, select the offered quality, and start the queued operation.
How does this agent compare with similar options?
Compared with yt-dlp alone, OmniGet adds graphical quality selection, queues, resume behavior, authenticated browser cookies, players, and post-processing, at the cost of installing a much broader desktop application. Compared with single-site web downloaders, it combines native course and social extractors, torrents, direct files, and roughly 1,800 yt-dlp-supported sites while retaining files locally. Compared with paid course downloaders, it is free under GPL-3.0 and covers several course platforms, although authenticated material still requires the user's legitimate session. On the agent side, it is a common desktop harness for Claude Code, Codex, Gemini CLI, ACP agents, and local models rather than a replacement for those underlying agents or model services.