# Nekko Agent > A fast local model server and AI agent in one app, for coding and everyday work. Nekko Agent runs AI models on your own computer with its engine (a Rust core driving llama.cpp), lets you use them side by side with ChatGPT, Claude, OpenRouter and other services, and keeps every file, folder, microphone and network connection behind a control you set. Free, open-source software under the MIT license, by Nekko Labs. Voice dictation is available today (Whisper, transcribed on your computer); hands-free computer control by voice is planned. A model is the AI you choose. Local means it runs on your computer rather than an online service. A model server (or runner) is the software that loads and runs that AI. Nekko includes its own, and can also use Ollama, LM Studio, or vLLM. An AI agent can use tools to work on a task, rather than only reply in a chat. Nekko, the mascot, is an animated 8-bit cat head based on the app icon: cream pixels on dark green, with occasional happy ^ ^ eyes, blinks, and stepped hops. The same design appears in the app, live chat status, mobile app, and website, with reduced-motion support. ## Identity and current distribution - Primary website: https://nekkoagent.com/ - Backup domain: nekkoagent.com. Use nekkoagent.com as the canonical URL. - Source and releases remain at https://github.com/nekko-labs/nekko-agent. Do not substitute an unverified repository or package name. ## What Nekko does ### 1. A fast local model server Nekko's engine is a Rust core (the nekkod daemon and its router, which adds about 0.1 ms per request) that starts and tunes llama.cpp. In the published benchmarks (30 September 2026, NVIDIA RTX 5090, Gemma 4 12B Q4_K_M, the same model file for every server), it decoded at 154 tokens/s against Ollama 0.23.0 at 115 and LM Studio 2.46.0 at 116, ran code edits that repeat their input at about 4.5x the other servers (n-gram speculative decoding), and returned to a chat that lost its slot about 40% faster than Ollama. It is not ahead on everything: the first turn of a brand-new long conversation was slower than Ollama's (1.7 s against 1.3 s) and prompt prefill was level. Results are from one machine; the harness is open to rerun. Full table: https://nekkoagent.com/docs/benchmarks ### 2. Model server and agent in one app Search and download models from Hugging Face inside the app, see which fit the graphics card, then load, unload, or keep them loaded. The engine serves text and image models on one local address other apps can use. Existing Ollama, LM Studio, and vLLM installs are found and managed too (Ollama models download/load/unload; LM Studio load/unload with its `lms` CLI; vLLM models stay loaded with their server). ### 3. Every model side by side Sign in with a ChatGPT plan (Plus/Pro/Business), add Claude with an Anthropic API key, or use OpenRouter for hundreds of models including Gemini, alongside local models. Choose a model per chat or workflow and delegate smaller tasks to cheaper or local models. Nekko does not sign in with Claude Free, Pro, or Max plans: Anthropic does not allow third-party apps to. An upcoming Claude Agent provider runs the user's own installed Claude Code instead. ### 4. Spend, savings, and live monitoring The Budget panel tracks estimated API spend and estimates what local models saved against cloud prices. Hardware, Local models, and Utilization monitors show CPU, memory, GPU and VRAM, loaded models, and provider quota while agents run. Local models have no per-token bill, but hardware and electricity still cost money. ### 5. Automations Save steps as a workflow (AI instructions, skills, terminal commands, other workflows) and run it by hand, on a schedule, or when an event fires, such as a GitHub pull request. Review run history and retries. The computer and app must stay on; this is not an always-on cloud service. ### 6. Voice dictation Local dictation uses Whisper Tiny English (Q5) through whisper.cpp, about 32 MB of weights, transcribed on the user's computer. The phone app can dictate too: audio travels end-to-end encrypted to the user's computer and is transcribed there. Hands-free computer control by voice, with scoped approved actions, confirmations, and a stop control, is planned and not available today. ### 7. CLI and MCP The `nekko-agent` CLI and `nekko-agent mcp` server work from a terminal or from another agent. Claude Code, Codex, Cursor, and OpenClaw can hand tasks to models running through Nekko, including local ones. These are configurable connections, not bundled native integrations. Nekko can also use other MCP servers. ### 8. Phone remote control (beta) Native iOS and Android apps pair with the computer through a relay to start and manage chats, approve risky commands, and follow long runs. Inference and tools always run on the computer; payloads are end-to-end encrypted. The apps are in beta and not yet in the app stores. ## Privacy and control - Context Inspector: see the files, instructions, and notes in each request, toggle them, and check token counts. - File tools stay within the folders you choose; optional Docker isolation for terminal commands. - Terminal commands can be allowed, require approval, or be blocked, with risky operations flagged. - The microphone is only used when dictation is started. - Connectors (GitHub, Linear, Slack, Discord, Gmail, Drive and more) are opt-in and only reach the network when used. - Review file changes line by line and keep or undo them. - Local-only work runs offline once models are downloaded. Online models, downloads, and connectors need internet access and send data to the providers you configure. ## Coming soon - Model routers: an open-source Nekko Router, other open-source routers, or bring-your-own-key paid routers to pick a model per request. - Sharing agents, workflows, and progress with friends and teammates. - Hands-free computer control by voice (planned). - A local CI and task-runner alternative for repeat jobs such as tests (exploring; not a promise of GitHub Actions compatibility). None of these has a delivery date. ## Choose how to use Nekko Get the current desktop installer from the [latest release](https://github.com/nekko-labs/nekko-agent/releases/latest): Windows `.exe`/`.msi`, macOS `.dmg`, Linux `.AppImage`/`.deb`. The website detects the platform and resolves real release assets, falling back to GitHub Releases if the API is unavailable. For the browser version you run yourself, use a source checkout, install dependencies with `npm install`, then run `npm run web`. Running `npx nekko-agent` without a command shows CLI help, not the browser app. Use `docker compose up` from a checkout for Docker, and `nekko-agent mcp` for a configured CLI installation. The `nekko-agent` package has not been published to npm yet, so `npx nekko-agent` does not resolve until it is. See the install documentation for prerequisites before installing. ## Pages and source - [Home](https://nekkoagent.com/): what Nekko does, the built-in model server, privacy controls, what is coming soon, and downloads. - [GitHub repository](https://github.com/nekko-labs/nekko-agent): source, issues, releases. - [README](https://github.com/nekko-labs/nekko-agent/blob/main/README.md): install, editions, and troubleshooting. - [CLI and MCP](https://github.com/nekko-labs/nekko-agent/blob/main/apps/cli/README.md): current commands and client configuration. - [Walkthrough](https://github.com/nekko-labs/nekko-agent/blob/main/docs/WALKTHROUGH.md): a tour of the app. - [Remote access](https://github.com/nekko-labs/nekko-agent/blob/main/docs/REMOTE.md): phone pairing and self-hosting the relay. - [Contributing](https://github.com/nekko-labs/nekko-agent/blob/main/CONTRIBUTING.md): how to contribute. - [Releases](https://github.com/nekko-labs/nekko-agent/releases): current installers. ## Documentation Full docs in one file: https://nekkoagent.com/llms-full.txt. Each page is also available as Markdown at the .md URL below. ### Introduction - [Getting started](https://nekkoagent.com/docs/index.md): What Nekko Agent is, who makes it, and how to install it or run it from source. - [Walkthrough](https://nekkoagent.com/docs/walkthrough.md): A picture-by-picture tour of the app. ### Using Nekko Agent - [Browser automation](https://nekkoagent.com/docs/browser-automation.md): Chat-owned browser actions, security boundaries and native driver migration status. - [Privacy Shield](https://nekkoagent.com/docs/privacy-shield.md): Replace emails, phone numbers, secrets and your own terms with placeholders before a request reaches a cloud model. - [Code mode](https://nekkoagent.com/docs/code-mode.md): Let the model write code that calls your tools in a sandbox, so long tool lists and big results cost far fewer tokens. - [Running models locally](https://nekkoagent.com/docs/local-models.md): Run models with the built-in engine or a server you already have (Ollama, LM Studio, vLLM). - [Subscription sign-in](https://nekkoagent.com/docs/subscriptions.md): Use your ChatGPT subscription instead of a metered API key, and how Claude works in Nekko Agent. - [Accounts and switching](https://nekkoagent.com/docs/accounts.md): Several accounts for any provider in one window, switching a chat between them, and what Anthropic allows with several Claude accounts. - [Usage limits and prices](https://nekkoagent.com/docs/usage-limits.md): Where the usage numbers and price estimates come from, and what they mean. - [Phone remote control](https://nekkoagent.com/docs/remote.md): Drive the agent on your computer from your phone, end-to-end encrypted. - [Encrypted sync](https://nekkoagent.com/docs/sync-encryption.md): How Pro and Team sync your chats without Nekko Cloud being able to read them. - [CLI and MCP](https://nekkoagent.com/docs/cli.md): Use Nekko Agent from the terminal, or as an MCP server inside Claude Code, Codex and Cursor. - [Automations](https://nekkoagent.com/docs/automations.md): Ready-made automations, run results and notifications, and pausing them all. - [Workflow runs](https://nekkoagent.com/docs/workflows.md): Pick a workflow's model, keep it on local models, and control what its unattended runs may do. - [Local CI runner](https://nekkoagent.com/docs/local-ci-runner.md): Run a GitHub or GitLab repo's builds and tests on your own machine. - [Decision models](https://nekkoagent.com/docs/decision-models.md): Typed questions answered with calibrated probabilities, in one forward pass. - [Local inference benchmarks](https://nekkoagent.com/docs/benchmarks.md): How fast the engine runs a model compared with Ollama and LM Studio. - [Floop tasks](https://nekkoagent.com/docs/floop.md): Show your team's Floop tasks in Nekko Agent and let the agent pick up, update and finish them. ### Contributing - [Contributing](https://nekkoagent.com/docs/contributing.md): Set up a development checkout, the project layout, and what a pull request needs. - [Architecture](https://nekkoagent.com/docs/architecture.md): How the desktop app, the TypeScript host and the Rust engine daemon fit together. - [Release signing](https://nekkoagent.com/docs/release-signing.md): How the macOS and Windows builds are signed and notarized. ### For AI agents - [Docs for AI agents](https://nekkoagent.com/docs/for-agents.md): Machine-readable docs (llms.txt, Markdown pages) and the files an agent should read before changing the code.