OpenYak Desktop Workbench
A local-first desktop agent for turning real files into structured work products with your chosen model boundary.
- Source repo
- openyak/openyak
- Stars
- ★ 693
- Last updated
- 2mo ago
- License
- Apache-2.0
- Primary language
- TypeScript
- FA score
- 63/100 · Some gaps
At a glance
- How it runs
- Works with
- Universal · cross-platformOpenAI API · Claude APIChatGPT (Partial support)
- Cost
- Free software; you pay for model usage
- Setup effort
- Medium · a few setup steps
- You'll need
- Typical use
- A department manager with a dense DOCX memo can turn it into an executive brief, risks, owners, next actions, and a send-ready follow-up email.
- Not a fit if
- Linux users needing native Computer Use: it ships only for macOS and Windows
- Users wanting zero setup: you must install a local model or supply your own API keys
- Teams that cannot send prompts and context to third-party cloud providers
- Source review
- 63/100 · Some gaps
What does this agent do, and when should you use it?
OpenYak is a local-first desktop agent workspace that keeps files, conversations, memory, workflow state, and generated artifacts on the user’s device. It works with DOCX, XLSX, PPTX, PDF, CSV, local folders, and generated artifacts, presenting results in chat and an Artifact workspace. The desktop product uses Tauri v2, Rust, and Next.js 15, backed by a FastAPI agent runtime and SQLite storage. Inference can run locally through Rapid-MLX, Ollama, or a custom OpenAI-compatible endpoint, or directly against user-configured cloud providers such as OpenAI, Anthropic, and OpenRouter. Its Ultra mode lets a parent agent create 2–4 durable child sessions and synthesize their results into the original task.
After a user installs the desktop app, they can start a conversation and attach working files. OpenYak reads office documents, spreadsheets, slide decks, PDFs, CSVs, local folders, and project context; with user-controlled permissions, it can read, write, rename, organize, and automate files, then produce Markdown briefs, tables, diagrams, checklists, and other structured artifacts. The Next.js frontend supplies chat, settings, an artifact panel, and SSE streaming, while the FastAPI backend handles the agent engine, tool execution, LLM streaming, and storage. For complex work, Ultra lets the parent agent launch AgentRuns in durable child Sessions, follow their state, exchange input, and synthesize persisted outputs. Model traffic goes either to localhost services such as Rapid-MLX, Ollama, or a custom OpenAI-compatible server, or directly to the cloud provider configured by the user.
- A department manager with a dense DOCX memo can turn it into an executive brief, risks, owners, next actions, and a send-ready follow-up email.
- A finance professional reviewing a budget workbook can request budget-versus-actual variance, drivers, anomalies, and meeting talking points.
- A project lead reconciling a memo, budget, slide deck, and PDF can create one board brief artifact in a single thread.
- An operations user who needs continuity can move from analysis to a RACI, 30-day plan, agenda, and follow-up drafts without restating the context.
- A user with a multi-part task can enable Ultra so a parent agent distributes focused work to 2–4 child sessions and combines the outputs.
- A user away from their desk can enable remote access, scan a QR code, and send work to their computer through Cloudflare Tunnel.
How do you install or deploy this agent?
Download the latest release for the target platform: macOS .dmg or .app, Windows x64 .exe installer, or Linux x64 .deb or .rpm. After installation, choose an inference location in Settings: Rapid-MLX on Apple Silicon macOS, Ollama on macOS, Windows, or Linux after downloading a model, or a custom OpenAI-compatible local endpoint. For cloud inference, provide your own provider API key; the ChatGPT option uses an existing Plus, Pro, Team, or Enterprise subscription when available. For source development, the only documented launch command is npm run dev:all, which starts the backend on port 8000 and frontend on port 3000; the supplied material does not document prerequisite Node.js, Rust, Python, or dependency-installation steps.
How do you use this agent?
Launch the app, start a new conversation, and attach a real file. Select a local Rapid-MLX or Ollama model, or a configured cloud provider, then ask for a deliverable such as three key takeaways, risks, owners, next actions, and a follow-up email based on the uploaded files. Review the chat response and the right-side Artifact panel, then continue in the same thread with requests for a RACI, plan, or revision. For complex work, enable Ultra and describe the goal normally; the parent agent decides whether to create child AgentRuns. For mobile access, enable remote access in Settings, scan the QR code, and connect through the mobile web client.
What are this agent's strengths and limitations?
- Files, conversations, memory, artifacts, and workflow state are stored locally, and local-model requests can remain on the machine.
- It handles office documents, spreadsheets, slide decks, PDFs, CSVs, and local folders, then turns work into reusable Markdown briefs, tables, diagrams, and checklists.
- It offers multiple model paths: Rapid-MLX, Ollama, custom OpenAI-compatible endpoints, and a broad set of BYOK cloud providers.
- The runtime explicitly includes persisted run state, resumable SSE streams, cancellation, bounded retries, tool argument validation, and layered permissions.
- Ultra keeps durable child sessions, their status, and synthesized output inside the parent task rather than requiring a separate batch form.
- Offline operation requires the user to install and prepare Rapid-MLX, Ollama, or another local OpenAI-compatible service and download a model.
- When a cloud model is selected, prompts and relevant context are sent directly from the desktop to that provider; local storage does not make cloud inference offline.
- File-writing, renaming, organization, and automation are governed by user-controlled permissions, so available actions depend on the approvals granted.
- Remote access depends on Cloudflare Tunnel and token-based authentication, making it a network-dependent feature.
- The documented source-development path provides npm run dev:all but not a complete list of prerequisite runtimes or dependency-installation steps.
How does this agent compare with similar options?
The README positions OpenYak closer to a local agent workbench than web chat products such as ChatGPT or Claude.ai: it centers local files, artifacts, tool permissions, and long-running work tied to the user’s machine.
Key facts side by side with the most closely related agents.
| Agent | Source review | Form / cost | Stars | Updated | Language | Full support on |
|---|---|---|---|---|---|---|
| OpenYak Desktop Workbench This agent | 63 · Some gaps | Desktop appFree + model costs | ★ 693 | 2mo ago | TypeScript | OpenAI API · Claude API |
| Milady Personal AI Assistant | 60 · Some gaps | CLIFree + model costs | ★ 356 | 4mo ago | TypeScript | OpenAI API · Claude API |
| MyAgents | 57 · Major gaps | Desktop appFree + model costs | ★ 591 | 4d ago | TypeScript | Codex · Claude Code · Claude API |
| OpenHuman | 51 · Major gaps | Desktop appFreemium | ★ 43k | today | Rust | — |
How does FollowAgents rate this agent?
Why each dimension lost points
Evidence shows layered permission policies (allow/ask/deny), per-app and per-origin approvals, user takeover and confirmation, local storage, and BYOK model choice. However, implementation details are not provided, and dependency security only mentions CI and signing without audit or vulnerability scanning evidence. Deductions: insufficient dependency security evidence, rollback and source attribution only partially mentioned.
README and CI show persisted run state, resumable SSE streams, cancellation, retries, child-run failure isolation, and multi-platform testing. But no specific failure message examples or error handling documentation. Deduction: insufficient failure message evidence.
Clearly targets desktop office scenarios, supports multiple file types and local/cloud models, provides Computer Use and Browser workspaces, and multi-platform support. But capability boundaries and trigger precision are not detailed. Deductions: insufficient description of capability boundaries and trigger precision.
Clear README, install notes, FAQ, known limitations (e.g., Linux Computer Use not in v1.5 release gate), Apache-2.0 license, and versioned release process. But no CHANGELOG file, and maintenance responsibility only via community links. Deductions: missing standalone CHANGELOG, unclear maintenance responsibility.
README shows various workflows and output examples (e.g., executive brief, budget analysis, artifacts), and emphasizes local-first and BYOK for cost control. But no performance benchmarks or user feedback. Deduction: lack of quantitative evidence.
README references multiple documents (e.g., AGENT_RUNTIME.md, EVALUATION.md) and CI tests, but no specific test results or external validation. Deduction: insufficient cross-source corroboration.
- Insufficient dependency security evidence: no dependency audit or vulnerability scanning results; check dependency manifests and CI security steps.
- Rollback mechanism only mentions persisted state, not how to restore or rollback to previous states; clarify.
- Source attribution only via community links, no clear maintainer identity; verify.
- Remote access feature involves Cloudflare Tunnel; ensure secure configuration.
Recent releases
Latest entries from the project's GitHub Releases. Notes are quoted from upstream as written, not rewritten or tested by FollowAgents.
-
Release notes excerpt
OpenYak 1.5 Release Candidate 3 RC3 makes the shared Computer and Browser workspaces behave like focused parts of the Agent task instead of floating UI. On wide screens, the live workspace now stays pinned beside the conversation, never covers Agent messages or the composer, and no longer stacks the compact Workspace summary underneath it. > Release candidate: this build is for hands-on testing before v1.5.0 stable. Stable users remain on v1.4.0 and will not be updated automatically to this RC. …View on GitHub ↗ -
Release notes excerpt
SHA-256 Checksums Verify your download by comparing its SHA-256 against the table below. On Linux: sha256sum <file>. On macOS: shasum -a 256 <file>. On Windows (PowerShell): Get-FileHash -Algorithm SHA256 <file>. | File | SHA-256 | Size | |------|---------|------| | OpenYak_1.4.0_aarch64.app.tar.gz | b7c78b58d9cb3f25c784da17f70d70d09856d246d14606425a829bd111fe87ed | 145.6 MB | | OpenYak_1.4.0_aarch64.dmg | 4fe97dbafba380ebd4c5573ef416ba43b10dd70426d59c3d8063313519f9f38a | 186.8 MB | …View on GitHub ↗ -
Release notes excerpt
OpenYak v1.3.0 Model selection is now per-session. Each chat remembers the model it was last using and restores it when you come back — so switching models in one conversation no longer silently changes every other one. Plus a fix for false "can't read images" warnings. What's New Per-session model memory The model picker in the top-left used to be a single global setting: change the model in one chat and every other chat changed too, so returning to an older conversation showed whatever model you last picked elsewhere — not the one that conversation was actually using. …View on GitHub ↗