Your personal AI workspace: choose the model, shape the workflow and see what you share.
Bring your preferred models, instructions and tools into one workspace. We’re building Neel Desktop so you can shape how your agent works—and inspect what it shares, what it does and what comes back.
Launch brief: The product is ready for a small pilot. The support owner is unassigned. The rollback plan is not documented.
03
Model call 2 1.7 s
Respond with the evidence
The next request includes the tool result. The model returns its answer; no files are changed.
1,803 in · 64 out 39.8 tokens/s
Context used on the final call 22%
1,803 / 8,192 tokens. Separate from total tokens across the task.
THE EXCHANGE
Exactly what goes to the model.
The instructions, your message, available tools and retrieved context—not just the last sentence in the chat.
System instructions
Summarise only the document the user selects. Ask permission before reading it. Report unresolved risks; do not change any files.
Your message
Summarise the launch brief and flag anything still unresolved.
Tool definition
read_document(path) · read a selected document after approval
Context returned by the tool
Launch brief: The product is ready for a small pilot. The support owner is unassigned. The rollback plan is not documented.
Inspect the complete second request
{
"model": "selected-local-model",
"stream": true,
"messages": [
{
"role": "system",
"content": "Summarise only the document the user selects. Ask permission before reading it. Report unresolved risks; do not change any files."
},
{
"role": "user",
"content": "Summarise the launch brief and flag anything still unresolved."
},
{
"role": "assistant",
"content": null,
"tool_calls": [
{
"id": "read_01",
"type": "function",
"function": {
"name": "read_document",
"arguments": "{\"path\":\"launch-brief.md\"}"
}
}
]
},
{
"role": "tool",
"tool_call_id": "read_01",
"content": "Launch brief: The product is ready for a small pilot. The support owner is unassigned. The rollback plan is not documented."
}
],
"tools": [
{
"type": "function",
"function": {
"name": "read_document",
"description": "Read a user-selected document after permission is granted.",
"parameters": {
"type": "object",
"properties": {
"path": {
"type": "string"
}
},
"required": [
"path"
]
}
}
}
]
}
Know the destination, too. This example runs locally. A hosted provider is an explicit choice; its data-handling terms still apply.
What came back. What you see.
Same answer. Nothing silently rewritten.
Before presentation
The raw model response
{
"role": "assistant",
"content": "The product is ready for a small pilot.\n\nBefore launch:\n- Assign a support owner.\n- Document the rollback plan.\n\nSource: launch-brief.md. No files were changed.",
"finish_reason": "stop",
"usage": {
"prompt_tokens": 1803,
"completion_tokens": 64
}
}
After presentation
The readable answer
The product is ready for a small pilot.
Before launch:
Assign a support owner.
Document the rollback plan.
Source: launch-brief.md. No files were changed.
Formatting only. No extra model call.
Inside the workspace
Power you can see and control.
The features below exist in the current development build. Early access is how they reach you first.
AI request details
Open any turn to see the prompt, the raw request and response, timing and tokens. Nothing about the exchange is hidden behind the chat.
Model settings
Choose the model, thinking effort, standard or fast speed, and context size, with memory needs shown before anything loads.
Mode and permissions
Plan or build, and per-tool permissions, sit right by the editor. Plan mode withholds every change; build mode works within the permissions you set.
MCP servers
Add and manage MCP servers from a sidebar, through a form or as JSON.
Terminal panel
An interactive terminal in a side panel, so commands and the agent share one window.
Attachments and goals
Attach files and screen captures, and set a goal the agent works toward.
Connections
Connect providers you opt into and pair a phone through Neel Mobile, which is in development.
Same engine as the CLI
Desktop and the CLI share one engine, one permission authority and one session store.
Models and hardware
Local models, sized to your hardware.
Explore Gemma, Qwen, Llama and DeepSeek, with curated defaults and bring-your-own options clearly separated. Choose a compatible local runtime or a supported provider you opt into.
Curated defaults
Gemma
A guided starting point
Neel’s curated Gemma 4 options come with documented memory guidance for the initial local runtime.
Gemma 4 E2B · E4B · 12B
Bring your own model
Qwen
Coding and multilingual work
A broad family for code, language and reasoning tasks, with smaller and larger variants to explore through a compatible local runtime.
These are model families to explore, not a list of Neel-tested builds. Beyond the curated defaults, bring a compatible GGUF model or connect a supported local server. Compatibility, tool use and memory needs depend on the exact model file, runtime and context length; not every variant is supported.
Curated Gemma memory guide
These documented defaults cover the Gemma 4 MLX 4-bit configuration on macOS with Apple silicon. This picker covers these three models only; it is not a cross-platform hardware guarantee.
Curated Gemma 4 MLX models and the free memory each needs
Model
Role
Free memory needed
Device memory
Gemma 4 E2B3.5 GiB on disk
Default
6.5 GiB
8 GB (tight)
Gemma 4 E4B
Comfortable
8.5 GiB
16 GB
Gemma 4 12B
Largest local
22.5 GiB
32 GB
These figures apply to the curated Gemma 4 MLX 4-bit variants on macOS, not every model family or runtime. Neel checks available memory before loading. Apple's on-device model offers another starting point where it is available, without a separate model download.
Privacy defaults
Nothing leaves without your say.
No account for local use
Local models run without signing in and without an inference bill.
Models download only when you choose
Neel ships without model weights. Nothing large downloads without your say.
Every download SHA-256 verified
Runtimes fetched on first use are pinned and checked against a SHA-256.
Uninstall keeps your data
Uninstalling removes only what we installed. Models, settings and conversations stay yours.
Status
Early access, honestly labelled.
Neel Desktop launches with the CLI, the editor extensions and Neel Mobile, planned for January 2027, on Windows, macOS and Linux. Every installer is signed, notarized on macOS and tested on a clean machine before a download link appears on this site.
Join early access and tell us which devices you use.
Status
Early access · In development · private workspace
Launch platforms
Windows, macOS, Linux
Launch
January 2027, with the whole ecosystem
Account
Not needed for local models
Package
Signed installers (in preparation)
Early access
Get Neel Desktop first.
Join the waitlist for early access, or email us about what you want to build with Neel. We will get in touch when it is ready for your machine.
