Describe the task, then stay in control
DJL turns a request into ordered steps, runs each tool in the open, and stops for your approval before anything touches your files.
Describe the task
Write what you want in English or 中文. Both languages share one context, so you can switch mid-task without repeating yourself.
Read the plan
Work is decomposed into ordered, legible steps you can read before anything runs.
Watch the tools
Each tool executes in a sandbox, streaming its output so nothing happens off-screen.
Review the diff
Every change waits for you: approve and apply, request changes, or cancel and roll back.
Nothing reaches your project until you approve it.
Run a model on your own hardware
DJL works out what your computer can run, then installs and starts the runtime itself. No terminal, no admin password.
Local model management lives in the DJL desktop app.
Open the desktop app
Download DJL for your platform and launch it.
Go to Settings → Local Models
Listed under Private inference.
Check what DJL detected
It reports your processor, graphics card, and memory, then recommends the largest model that will still feel fast — deliberately one tier below what your machine could merely hold.
Install and start a runtime
Choose Ollama or LM Studio. DJL downloads it, starts the local server, and connects to it.
Select the model
Once installed, pick it in the model picker and describe your first task.
Two runtimes
- Ollama
- DJL installs, starts, and connects Ollama for you. No terminal or admin password needed.
- LM Studio
- Install LM Studio once, then DJL can start and manage its local server.
Curated models
DJL recommends one of these from your hardware. You can always choose another.
| Model | Memory | Download | Drives the agent |
|---|---|---|---|
| Qwen3 1.7B | 4 GB | 1.4 GB | Chat only |
| Qwen3.5 2B | 8 GB | 1.9 GB | Chat only |
| Granite 4.1 3B | 8 GB | 2.1 GB | Yes |
| Qwen2.5 Coder 7B | 16 GB | 4.36 GB | Yes |
| GPT-OSS 20B | 16 GB | 13 GB | Yes |
| Qwen3 Coder 30B | 32 GB | 19 GB | Yes |
“Chat only” means chat only
Qwen3 1.7B and Qwen3.5 2B answer questions well but are too small to hold a tool-calling loop together. Give either one a task that needs file edits and it stalls silently instead of reporting an error. For real work pick a model marked “Drives the agent” — Granite 4.1 3B is the smallest that qualifies, and it runs in 8 GB.
If the tools stop working
A capable model still needs enough context loaded to use tools. When DJL says the model is chat-only at its current context, reload it with a larger context window and the tools come back.
Bringing your own model
The curated list is a starting point, not a limit. Install any Ollama tag, LM Studio catalog ID, or exact Hugging Face model URL.
What stays on your machine
Inference runs locally: prompts, code context, and output are never sent to a hosted model provider. DJL connects only to fixed loopback addresses (127.0.0.1) and never exposes either runtime to your network. Tools you approve can still reach the network on their own.
What stays on your machine
Inference runs locally: prompts, code context, and output are never sent to a hosted model provider. DJL connects only to fixed loopback addresses (127.0.0.1) and never exposes either runtime to your network. Tools you approve can still reach the network on their own.
Download DJL and give it a model that runs where you are.
Download for macOS