NEKKO AGENT
Download
FREE & OPEN SOURCE · RUNS ON YOUR COMPUTER
Nekko Agent
The personal, local agent
you can trust.
Cloud models are great. Would you trust them with your data and even your wallet? Nekko was built around security and providing a high quality experience without compromising your data.
No account needed to use AI on your computer · Works offline after setup · MIT licensed
An agent built for local models. Finally.
Private, fast, and on every machine you own.
-
No separate model engine to install: it is built in. Claude Code, Codex and Cursor can use your local models through the CLI and MCP server.
CLI and MCP -
A whole-file code edit in 2.5 s against 13.9 s on Ollama: same model, same machine. Speculative decoding helps most when an edit repeats its input.
Benchmark summary -
Local chats never leave your computer. Privacy Shield hides personal details from cloud models, and your phone reaches your computer end-to-end encrypted. Even we can’t see in.
How Privacy Shield works -
Cloud models, local models, and that 5090 and Mac mini in the corner. Connect your computers and pick any of their models from one place.
Connecting computers -
macOS, Windows and Linux, plus iPhone, Android, iPad and folding phones. Approve a command from your pocket while the work runs at home.
Phone apps are in beta: get early access
Benchmark summary
- Whole-file rename benchmark
- 2.5 s vs 13.9 s on Ollama 0.40.11
- Apple M5 Max · Qwen3-8B, same GGUF weights · cool-chip medians, 8 requests per server · October 2026
- Code edit
- 534 vs 106 tokens/s on Ollama
- NVIDIA RTX 5090 · Gemma 4 12B · Ollama 0.23.0 · 30 September 2026
- Four requests at once
- 256 vs 113 tokens/s on Ollama, 210 on LM Studio
- NVIDIA RTX 5090 · Gemma 4 12B · LM Studio 2.46.0 · 30 September 2026
Where we are not ahead: MLX decodes plain text faster on Apple Silicon, speculation adds first-token overhead, and Ollama starts a fresh long chat faster. These time inference replies, not a complete agent workflow. Full method, raw runs and limitations
Also in the box: scheduled agents and triggers, Whisper dictation on your computer, cost tracking for cloud models, skills and sub-agents, and connectors for GitHub, Linear, Slack, Discord, Gmail and Drive. The demos are illustrations in the app’s design; the speed demo replays published benchmark medians.
More is coming.
Want to get exclusive early features? Join our beta testers list.
Want to try it first?
- A chance to try the iOS and Android apps before anyone else
- First to test new features as they land
- More beta-only extras to come
Psst… want to see what's next?
We'll only use your email (and name, if you give it) to send beta invites, and product updates only if you tick the box. No spam, and we never share it. To be removed, email hello@nekkoagent.com. Privacy policy.
Your own model server, built in
Nekko comes with its own engine, so running AI on your computer takes a few clicks rather than a weekend of setup. Find a model, see whether it fits your machine, and start using it.
- Search and download models from Hugging Face without leaving the app
- See which models fit your graphics card before you download them
- Load, unload, or keep a model ready so your first reply comes back fast
- Serve text and image models on one address that your other apps can use too
- Already using Ollama, LM Studio or vLLM? Nekko finds them and manages their models
- Add ChatGPT, Claude, OpenRouter or any OpenAI-compatible service alongside
Example setup, not a scan of your machine
built inlocalhost:11434localhost:1234Share only what you choose
Letting an AI read a prompt is one thing. Handing it your files, your terminal, your microphone and your network is another. Nekko keeps every one of those behind a control you set, and shows you exactly what it did.
Choose the information it uses
Context means the information sent with your request. Check what is included and leave out what is not needed.
- Context Inspector See the files, instructions, and saved notes included in your next request. Toggle or pin them and see how much space they take up, measured in tokens, the small pieces of text AI models process.
- Find relevant files Search across the project folders you choose, so Nekko can find useful code and notes without sending every file to the AI.
-
Guidelines and memory
Give Nekko project instructions through
AGENTS.mdorCLAUDE.md, and keep saved notes for your preferences. Turn individual instructions and notes on or off. -
Save the plan
Turn the conversation into a project description in
SPEC.md, plus a plan and task list that can guide later work. - Improve your request Get suggestions for making your instructions clearer before you send them. This draft check runs on your computer without an online AI service.
Give it useful tools
Connect the tools and repeatable steps your work needs, rather than copying everything into a chat.
- Split up a task Delegate smaller pieces of work to separate AI sessions, called sub-agents. Each gets its own instructions, with limits on further delegation and a view of the work in progress.
- Reusable instructions Not sure which skill to reach for? Nekko's library sorts skills by the job they do, from research and planning to code review and testing, and shows every step before you run one. Install more from the marketplace in a click.
- Connect coding tools MCP is the common language AI tools use to talk to each other. Nekko can use other MCP servers or act as one, so Claude Code, Codex, Cursor or OpenClaw can pass work to your local models through it.
- Connect your services Bring in what matters from GitHub, Linear, Slack, Discord, Gmail and Drive instead of pasting it by hand. Each connection is one you switch on, and it only reaches the network when you use it.
- Longer tasks with limits Give Nekko a goal to plan and work through. Set limits on time or steps, and get a request for guidance if it stops making progress.
Review and correct changes
AI can make mistakes. Approval settings, file limits, and change review help you stay involved.
- Command approvals Choose which terminal commands are allowed, need approval, or are blocked. The app checks for risky operations such as deleting files or changing system permissions.
- Limit file access Nekko only touches the folders you pick. Your microphone is only used when you start dictating, and speech is turned into text on your computer. For extra peace of mind, Docker can wall off terminal commands entirely.
- Review file changes See edits line by line in the Changes panel. Keep or undo individual lines, whole files, or all changes from the app.
- Try again from an earlier step If an AI returns an empty reply, Nekko retries and then reports the problem. Use "Reset here" to return to an earlier message and try a different approach.
Meet Nekko: a small pixel cat keeping you company while your agents get to work.
Choose what stays on your computer
Work without an internet connection
Once your model is downloaded, local work keeps going without the internet. Read, edit and search your folders with nothing sent to an online AI service.
Connect online AI when you need it
Online AI models, downloads, and connected services need internet access and may send data to the providers you configure. You choose those connections and are responsible for their costs.
Choose how to use Nekko
Download the desktop app to get started. If you prefer a technical setup, run Nekko in your browser, in Docker, or from a terminal.
In your browser
Run it yourself from a source checkout: npm install, then npm run web. Tasks using AI on your computer can work offline after setup.
In Docker
For container-based setups, run docker compose up from a source checkout and connect your project folders and model servers.
From other tools
The CLI is Nekko's command-line interface. Run nekko-agent, or nekko-agent mcp to connect tools that support the MCP standard.
The npm package is nekko-agent, publishing with the next release; until then run the CLI from a source checkout (npm run dev or apps/cli). It provides the nekko-agent command.
What we are building next
Here is what is on the bench right now. Join the beta to try each one first.
Model routers
Let a router pick the right model for each request, so easy questions go somewhere cheap and hard ones get the heavy hitters. Use our upcoming open-source Nekko Router, another open-source router, or bring your own key for a paid one.
Share with friends and teammates
Great work is rarely a solo act. Share agents, workflows and progress with the people you build with, without giving up control of your own machine.
Control your computer by voice
Dictation is here today. Next, Nekko listens, acts and speaks back, all on your computer: you approve which apps and actions it can use, confirm sensitive steps, and always have a stop button in reach.
Run project checks on your machines
A local alternative to hosted CI for repeat jobs such as running tests, with the option of cloud runners for extra capacity. This is not a promise of GitHub Actions compatibility.
Download Nekko Agent
Free, open-source software. Built by an investor-free team, independent from any external influences.
Windows builds are unsigned for now (SmartScreen: More info → Run anyway); free code signing by the SignPath Foundation is in progress.
OPEN SOURCE. BUILT TOGETHER.
The people behind Nekko.
Our Nekko Labs team and community is amazing! Huge thank you to the people sharing their time, ideas and code to make Nekko Agent yours.
Meet the contributors on GitHub.