Draggy
Chat with a local model
On first launch Draggy sizes a model to your graphics card and downloads it. Chat and Code are kept apart, each with its own history, model and settings.
A desktop assistant that runs on your own computer.
Draggy talks, browses and works with your files using models on your own machine. It drives a local Ollama, so there is no account and no API key.
Free software under the GPL. Windows, macOS and Linux. Version 1.2.7.
No account
Nothing to sign up for and no key to paste.
No telemetry
None is collected, and there is nowhere to send it.
Your machine
Chats and documents stay on your own disk.
GPL v3
Free to use, read, change and pass on.
What it does
An assistant that can actually do things
Answering questions is the easy half. Draggy also reads your files, writes new ones, searches, and opens pages for itself.
Works in your projects
Switch to Code and open a folder. Draggy reads, edits, moves and deletes files in it, and only in it. Every change shows as a diff and can be undone.
Asks before it acts
Each project gets a mode, from only proposing to working on its own. Anything outside it stops and asks, with the change or the command shown first.
Plans you can steer
The model writes its plan as a checklist you can tick, reorder and edit while it works. Tasks carry on in the background.
Writes real files
Word, PowerPoint, Excel, PDF, code and plain text, saved where you can open them. Attach a document and it reads that back.
Looks things up
It can search the web, open a result and read it, or drive a real browser session when a page needs clicking through.
Searches your documents
Point it at a folder. The contents are indexed on your own machine and searched on meaning and keywords at once.
Runs commands
In Code it runs git, tests and builds in the project folder, and short Python and JavaScript in a sandbox. A command asks before it runs.
Extends with MCP
Forty-five MCP servers in a catalogue, the official registry to search, and remote servers with sign-in, all off until you turn them on. Plus 64 skills that teach it how a job is done.
Talks back
A continuous voice mode that works out when you have finished a sentence, answers out loud, and stops when you cut in.
Twelve languages
English, French, Spanish, German, Italian, Portuguese, Dutch, Russian, Chinese, Japanese, Korean and Arabic.
Keeps itself current
Updates download in the background and install the next time you open the app. Nothing is installed behind your back.
Projects
Point it at a folder
A project gives Draggy one folder of yours to work in. It follows the AGENTS.md you keep there, shows every edit as a diff, and keeps the old version so any change can be undone.
- Open a file beside the chat and edit it together.
- See the branch and what changed, straight from git.
- Undo any change, whether or not the folder uses git.
Commands
It asks before it runs anything
In Code the model runs commands in your project: git, gh, package managers, tests and builds, in PowerShell, cmd, bash or sh. You see each command before it runs, and nothing runs until you allow it.
- Allow it once, for the task, or always.
- Always allowing remembers how a command starts, never what is chained after it.
- Stopping a turn stops the command and everything it started.
Extensions
Plug in what you already use
Switch on MCP servers from a catalogue of forty-five, search the official registry for more, or add a remote server with sign-in. Skills teach it how a job is done: 64 come with Draggy, and you can write your own.
- GitHub, GitLab, Sentry, databases and more, each off until you turn it on.
- Type / and a skill's name to start it, or let the model pick the one that fits.
- A server can answer with a small interface of its own, kept in a sandbox.
Voice mode
Say it instead
Voice mode listens continuously and works out when your sentence has ended, so there is no button to hold and no wake word to remember.
- It answers out loud, and stops the moment you cut in.
- Speech recognition and the voice both run on your machine.
- Captions on screen for everything either of you said.
Built-in browser
Pages without the adverts
Links open in a browser window inside the app, running uBlock Origin's filter lists on Ghostery's engine. The model reads the same clean page you do.
- YouTube adverts and anti-adblock walls included.
- Browsing runs in its own session, kept away from the app.
- Switch the blocker off per window when a site needs it.
History
Conversations you can take with you
Every chat and every Code session is stored on your own disk and searchable from its history list. Any chat exports to a Markdown file in a click.
- Search across everything you have ever asked.
- Export to Markdown, which every other tool can read.
- A long conversation is condensed rather than dropped.
Statistics
See what every reply costs
Every finished reply adds to a record kept only on this machine: speed, tokens, which tools ran, and how long tasks actually take. Nothing is sent anywhere to measure it.
- Speed and time to the first token, broken down by model.
- Which tools get called, and how often.
- How long a task takes: the average, and the length nine in ten finish under.
Requirements
What it needs to run
The model runs on your hardware, so your hardware decides how fast it answers. Ollama has to be installed; Draggy offers to install it if it cannot find it.
Minimum
It will run| System | Windows 10 64-bit (1809 or newer), macOS 11, 64-bit Linux with glibc 2.28 or newer |
| Processor | Intel Core i5-8250U, AMD Ryzen 3 3200U, Apple M1 |
| Memory | 8 GB |
| Graphics | Intel UHD 620, AMD Radeon Vega 8 |
| Disk | 10 GB free |
Recommended
It will feel quick| System | Windows 11, macOS 14, a current Linux |
| Processor | Intel Core i5-12400, AMD Ryzen 5 5600, Apple M2 |
| Memory | 16 GB |
| Graphics | GeForce RTX 3070, Radeon RX 7600, Intel Arc A750 |
| Disk | 20 GB free, on an SSD |
Privacy
What actually leaves the machine
Model downloads, searches you or the model trigger, pages the browser opens, the speech models on first use, the ad blocker's filter lists, site icons for search results, the update check, a registry search when you type one, and any extension you switch on, remote ones included. Everything else stays local, statistics included. There is no telemetry, and nowhere for it to go.