Dulus
A desktop AI agent made of many agents — it drives your computer, follows you to your phone, and runs on whichever model you pick.
- Source repo
- KevRojo/Dulus
- Stars
- ★ 906
- Last updated
- today
- License
- GPL-3.0
- Primary language
- Python
- FA score
- 53/100 · Major gaps
At a glance
- How it runs
- Works with
- Universal · cross-platformOpenAI APIClaude API (Partial support)
- Cost
- Free tier plus a paid hosted plan
- Setup effort
- Low · running in minutes
- You'll need
- Typical use
- Office workers who want an AI to fill forms, click buttons and organize files inside real desktop software, not just chat.
- Not a fit if
- Teams that require every component to be open and auditable
- Developers who want a pure code library, not a desktop GUI
- Compliance contexts where no data may leave the machine
- Website
- dulus.ai ↗ set on GitHub
- Maintenance
- Commits in the last 30 dayslast commit 2026-11-20latest release v7.0.2 (2026-11-19)
- Source review
- 53/100 · Major gaps
What does this agent do, and when should you use it?
Dulus is an AI agent for the desktop and the terminal. The GitHub repository KevRojo/Dulus is open source under GPLv3 and ships a terminal CLI installable with pip install dulus; the fuller Dulus desktop app is built on a private codebase and downloaded from dulus.ai/download, bundling the GUI, CLI, browser pane, voice and a Telegram mobile path. Execution is model-agnostic: 40+ provider integrations (major APIs, local runtimes such as Ollama, LM Studio and llama.cpp, plus Dulus-hosted models) sit under one reasoning dial, while the agent reads and writes files, runs shell commands, and drives screens and browsers. A mechanism called the Round Table (Mesa Redonda) seats up to 26 agents of mixed models at a shared task board that split work, call each other by name, and surface decisions as cards with options and a recommended pick. Extensions cover roughly 1,800 MCP servers, Skills, live-reloading plugins and AutoAdapter, which turns a GitHub repo into verified tools. Pricing is freemium: free with your own keys or local models, with paid Fuel plans for hosted power.
Dulus reads your files and shell environment, calls the model provider you selected (or a local Ollama/LM Studio/llama.cpp runtime), then acts: in the desktop app it captures the screen, clicks and types into any app, navigates a real browser pane and runs commands in a live terminal panel. The Round Table seats up to 26 agents of different models on one shared task board where they split work and call on each other by name, returning decisions as cards with options and a recommended pick. Memory stays attached to the profile and project it came from; stale facts get flagged, newer facts replace older ones, and old ones are archived rather than deleted, with compacted conversations remaining searchable. An Artifacts library versions documents and mini apps the agent builds, while Telegram pairing via QR code lets the same conversation continue on a phone with text, voice notes, photos and files, and approvals as buttons. A cache-aware request shape keeps repeated prompts served from provider caches; the README reports 91.9% of input tokens served from cache.
- Office workers who want an AI to fill forms, click buttons and organize files inside real desktop software, not just chat.
- Developers who want a terminal assistant that reads and writes a repo, runs commands and connects MCP servers via the CLI.
- Teams that need several models' conclusions side by side, using the Round Table to split work across up to 26 agents and collect decision cards.
- Frequent travelers who pair Telegram by QR and continue the same desktop conversation from a phone, approving actions with buttons.
- Individuals who want to keep cost and privacy in check by running local models through Ollama, LM Studio or llama.cpp.
- Content creators generating images and video with their own keys (FLUX.2, GPT Image, Veo, Sora 2 and more) and feeding results back in as references.
Key setup facts
Pulled from the install and usage notes below. The project README has the current values.
- Install / run
pip install dulusdocker pull ghcr.io/kevrojo/dulus
- Runtime
- Python 3.10+
- Install notes cover
- Windows · macOS · Linux · Docker
How do you install or deploy this agent?
The terminal CLI needs Python 3.10+ and installs straight from PyPI:
pip install dulusThe desktop app (Windows 10/11 64-bit, v6.4.4, per-user installer with embedded Python) comes from the download page; the README says macOS and Linux builds publish on the same page:
https://dulus.ai/download/Windows users can also run the published container image:
docker pull ghcr.io/kevrojo/dulusA free Dulus account is required only for hosted models, Fuel, Dulus Online and Dulus OS; using your own keys or local models needs no payment. Running local models first requires installing Ollama, LM Studio or llama.cpp yourself.
How do you use this agent?
After installing, pick a model provider in configuration: enter API keys for the major APIs, point at a local Ollama / LM Studio / llama.cpp endpoint, or use a dulus_sk_live_… key for Dulus-hosted models. Then issue a task in the desktop app or terminal; Dulus reads and writes files, executes shell commands and drives the screen and browser, showing a live overlay so you can watch, approve or stop it. For multi-agent work, open the Round Table and seat up to 26 models on one task board. On mobile, pair Telegram by scanning a QR code inside the app, then continue with text, voice notes, photos and files while progress and approvals arrive as messages and buttons.
Account and Fuel management: https://dulus.online/account/#plans
Fuel top-up: SOL or $DULUS on Solana, 1 Fuel = 1 $DULUSWhat are this agent's strengths and limitations?
- Broad model coverage: the README claims 40+ provider integrations plus local runtimes (Ollama, LM Studio, llama.cpp), which limits lock-in to any one vendor.
- Multi-agent coordination is a first-class feature: the Round Table holds up to 26 seats sharing a task board and returns decision cards with recommended picks.
- Cache-aware request shaping: the README reports 1,783.8M input tokens processed Nov 1–15, 2026 with 91.9% served from cache, and cached tokens are free in Fuel.
- Substantial test surface: 7,402 automated tests are documented, with the latest full run (Nov 8, 2026) passing 7,362 and 24 known failures tracked.
- Multiple delivery shapes share one account: desktop app, CLI, browser pane, Telegram mobile path and Dulus OS.
- Open core, closed product: the desktop app is built on a private codebase the README explicitly says is not published, so its full implementation cannot be audited.
- Vendor-tied services: hosted models, Fuel, Dulus Online and Dulus OS all require a Dulus account and dulus.ai services.
- Screen control and voice are app-only: there is no evidence in the repository that clicking, typing, Astral Voice or Artifacts exist in the open-source CLI.
- Local inference has real prerequisites: running local models means installing and maintaining Ollama, LM Studio or llama.cpp plus model weights.
- Some platform details are left open: the README says macOS and Linux builds publish on the download page but gives no specific versions or commands.
How does this agent compare with similar options?
Key facts side by side with the most closely related agents.
| Agent | Source review | Form / cost | Stars | Updated | Language | Full support on |
|---|---|---|---|---|---|---|
| Dulus This agent | 53 · Major gaps | Desktop appFreemium | ★ 906 | today | Python | OpenAI API |
| Ankita Desktop AI Workspace | 69 · Some gaps | Desktop appFree + model costs | ★ 15 | 7d ago | JavaScript | ChatGPT · OpenAI API |
| Skales | 55 · Major gaps | Desktop appFreemium | ★ 2.2k | 11d ago | — | ChatGPT · OpenAI API · Claude API |
| TuriX | 41 · Major gaps | CLIFree + model costs | ★ 3.2k | 2mo ago | Python | Claude.ai |
How does FollowAgents rate this agent?
Why each dimension lost points
SECURITY.md documents a tiered permission model (auto/manual/plan/accept-all), a safe-command whitelist and a mutation gate, so least_privilege and user_confirmation are substantively designed; data_flow_transparency notes prompts go to the chosen provider and plugins/MCP may call out, which is adequate. Sensitive data handling is only XOR+base64 obfuscation that the doc itself calls non-cryptographic, so it is deducted; the dependency surface is very wide (composio, posthog, sentry, litellm, playwright, PyQt6) with no lockfile or audit evidence, so dependency_security is 1; external effects (browser/screen control, Telegram, plugin execution) are described but lack fine-grained constraints; rollback has checkpoint/undo/fork plus an e2e test, scoring 2; source attribution is README self-assertion only, scoring 1.
agent.py handles interrupted streams, empty turns and error rollback with clear comments, and failure_messages has last_error/last_reply, scoring 2; however README, pyproject and FAQ contradict each other on version (v6.4.4 vs 3.15.4), test counts (7,402 vs 5,734) and provider counts (40+ vs 11+100), so self_consistency is 1; dependency_availability is 1 because pyproject and requirements.txt disagree (mempalace both core and optional) and many heavy deps install by default.
Covers terminal/desktop/Web/Telegram scenarios with a broad audience; capability boundaries (closed desktop app vs open CLI) are stated; SKILL.md descriptions have precise triggers (defuddle explicitly excludes .md); environment fit spans Windows/macOS/Linux/Termux/Docker, scoring 2.
Install docs are very thorough (four methods, troubleshooting), and examples_and_faq plus license (full GPL-3.0 text and metadata) earn full marks; information architecture is clear; but version numbers disagree across README, pyproject and GETTING_STARTED, so naming_stability and versioning_changelog are 1; maintenance responsibility is a solo developer plus AI agents with an unverified publisher, scoring 1; known_limitations are listed in SECURITY.md, scoring 2.
Output usability has REPL/WebChat/Telegram channels and structured events, scoring 2; but README carries heavy marketing claims (100% cache hit, 3x Fuel, token economics) that diverge from the open core's actual capability, so marginal_value and cost_benefit are 1 because the core value is diluted by commercial narrative and default installs pull many non-essential heavy deps.
README claims numbers come from a repo snapshot and dashboard, but no raw data file in the revision can be checked, so claim_traceability is 1; cross_source_corroboration is 1 because README, FAQ and pyproject contradict each other; fact_inference_separation is 1 because claims like '100% cache hit rate' conflict with the body's 91.9%.
- README marketing numbers (100% cache hit, 7,402 tests, 40+ providers) contradict the body, pyproject and FAQ and should not be treated as reliability evidence.
- API keys are only XOR+base64 obfuscated, which the docs admit is not cryptographic; use OS permissions or an external secret manager.
- Default install pulls many heavy deps (posthog, sentry, composio, playwright, PyQt6) with no lockfile; assess supply-chain risk yourself.
- accept-all mode removes the approval boundary, and plugins/MCP/skills run with user permissions; review sources before enabling.
- The desktop app is closed-source, so capabilities described on the README page cannot be verified from this repository's source.
- Publisher identity is unverified by FollowAgents and maintenance rests on a single developer.
Recent releases
Latest entries from the project's GitHub Releases. Notes are quoted from upstream as written, not rewritten or tested by FollowAgents.
-
v7.0.2 Dulus v7.0.2 Desktop app (private repo)
Release notes excerpt
Dulus v7.0.2 Nuevo mercado de plugins con ~1,500 integraciones, conexiones MCP con inicio de sesión OAuth y soporte SSE, apertura de cualquier carpeta como proyecto, WebDrive embebido sin robar foco, paquete .deb funcional para Linux y un paquete amplio de mejoras de seguridad y temas en toda la interfaz. 🧩 Plugins Market y gestión de plugins - Nuevo Plugins Market: explora ~1,500 plugins desde la app (junto a Skills y MCP), ordenados de la A a la Z, con filtros por categoría y etiquetas, enlace directo al repositorio y botón de Auto-adapt en un clic. …View on GitHub ↗ -
v7.0.1 Dulus v7.0.1 Desktop app (private repo)
Release notes excerpt
☁️ Websites: dulus.ai ya vende planes y enseña la factura real Planes a la venta: /pricing/ + sección en el home venden Operator $49 y Company $99 al mes. Cada pago vuelve como Fuel x3 ($147 y $297). Todo botón va a dulus.online/account/#plans con checkout Dodo amarrado al email logueado; sin login dice que el plan está a un sign-in de distancia. (c513303d) Factura real, no promesa: tarjeta "On a real bill" + proof del home con el dashboard del proveedor 1–15 nov 2026: 1,783.8M input tokens, 91.9% desde caché. Reemplaza el 70.4% de 155.7M de octubre. (c513303d) …View on GitHub ↗ -
v6.4.3 Dulus 6.4.3
Release notes excerpt
Dulus v6.4.3 — desktop builds for Windows and macOS. - Fresh Windows installer and macOS app bundle (6.4.3), smoke-tested before packaging. - Windows: embedded Python, every library and the desktop GUI. No Python install needed. - macOS: universal app bundle with embedded Python and every library. No Python install needed. Requires: Windows 10 or later (64-bit) · macOS 13 Ventura or later (Apple Silicon).View on GitHub ↗
FAQ
How far does the free tier go?
What credentials or runtime do I need before the first run?
Can it act on my computer without my consent?
Is $DULUS an investment?
Where do bug reports go?
Is Dulus open source?
Is Dulus safe to use?
Is Dulus still maintained?
Who makes Dulus?
What is Dulus's official website?
Thanks — we use this to improve the page.