Desktop agent
Chat with models, search the web, work with files, and create documents from the same app that runs and shares inference capacity.
Request a demoWhat it does
The Aquaduck desktop agent lets you choose a model, where it should run, and which tools it should use. Chat with it, ask it to research the web, analyze uploaded files, or create documents and PowerPoint slides.
Why it matters
Work stays close to your data by default. You can start with local inference, use capacity across your private network, or choose an approved cloud model without moving to a different application.
Deep dive
The Aquaduck desktop agent is a workspace for AI. It can hold a conversation, work through a multi‑step task, use selected tools, retrieve information from uploaded files, and produce finished documents. The agent lives in the same desktop app as the inference runtime, so one download gives you a place to work with an agent and a way to make your device’s capacity available to the network.
For each agent workflow, you can decide which model to use, where inference should run, and which tools the agent may access. That keeps the interface consistent while letting the execution path change with the task.
Choose the model, environment, and tools
An agent workflow combines three independent choices:
| Choice | What it controls | Example |
|---|---|---|
| Model | The intelligence, speed, context length, and capabilities available to the task | Choose a fast local model for chat or a stronger model for complex analysis |
| Execution environment | Where model inference runs | Run on‑device, across approved machines on the private network, or in the cloud |
| Tools | What information and actions the agent can access | Enable web search, browsing, file retrieval, or document creation |
From chatbot to working agent
A model is trained on lots of information, so you can ask questions, develop ideas, draft text, or continue a conversation. Giving an agent tools means it can gather new information for the underlying model and produce useful results.
| Capability | What the agent can do |
|---|---|
| Web search | Find current sources and information beyond the model’s existing knowledge |
| Web browsing | Open pages, follow links, extract details, and work across websites |
| File upload and RAG chat | Retrieve relevant passages from uploaded files and use them to answer questions |
| Document analysis | Summarize, compare, extract, classify, and reason over documents |
| Document creation | Draft structured reports, briefs, plans, and other files |
| PowerPoint creation | Turn research, source material, or an outline into a presentation |
| Tool selection | Use the tools appropriate for a particular task |
Work with your own files
File upload makes the agent useful for work grounded in information you already have. With retrieval‑augmented generation (RAG), the agent searches the uploaded material for passages relevant to the current question and adds those passages to the model’s context. This is more useful than asking a model to rely on general knowledge when the answer should come from a contract, report, policy, dataset description, or internal brief.
You can use that workflow to:
- ask questions about one document or a collection of files
- compare agreements, reports, or proposals
- extract requirements, risks, dates, or decisions
- turn source material into a summary or structured report
- convert research and documents into a PowerPoint presentation
Select where the work runs
The execution environment can match the privacy, capacity, and capability requirements of the task. The agent remains the same across execution environments.
| Environment | When to use it |
|---|---|
| Local | Keep inference on the current device for private, low‑latency, or offline work |
| Private network | Use approved devices for larger models, pooled memory, or additional capacity while keeping inference within the organization’s network |
| Cloud | Use an approved external model for complex or specialized tasks when policy permits |
Example agent workflows
| Workflow | Suggested setup |
|---|---|
| Review a confidential agreement | Upload the file, enable retrieval, and use a local or private‑network model |
| Research a market | Enable search and browsing, then ask the agent to synthesize findings into a report |
| Build a presentation | Combine uploaded source material with document and PowerPoint creation tools |
| Analyze a large document set | Use RAG with a model running locally or on higher‑capacity private devices |
| Solve a complex non‑sensitive problem | Select an approved cloud model while keeping the workflow in the desktop agent |
| Everyday private chat | Choose a local model with no external tools enabled |
Agents beyond text
AI agents can work with more than text. Depending on the model and tools available, an agent may accept audio, video, or images as input and produce those media types as output. That can support workflows such as transcribing and summarizing a meeting, inspecting an image, extracting events from video, generating a visual, or combining several media types into one task.
One download, one workspace
Aquaduck puts the desktop agent and inference runtime in one application. You can use the agent, manage models, select tools and execution environments, and make eligible device capacity available to your private network from the same interface.