Skip to main content
PromptQuorumBuilt for humans. Structured for AI.
Home/Power Local LLM
Mobile & Edge LLMs

Kai 9000 Review: Free Open-Source AI Assistant, Cloud or On-Device

·8 min readBy Hans Kuepper · Founder of PromptQuorum · Discovery engine for open-weight & open-source AI

Kai 9000 (version 3.3.0, released 2 October 2026) is an open-source assistant with persistent memory that only runs locally if you download an on-device model. By default it sends your messages to a cloud service. PromptQuorum has not installed or tested it; every statement here comes from the project's own sources.

Key Takeaways

  • Maintenance: GitHub shows release 3.3.0 on 2 October 2026 and a code push on 24 October 2026, so no stale-project warning applies.
  • Cost and licence: free under Apache 2.0, with no account; cloud providers you add charge you directly.
  • Reach: Android, iOS, Windows, macOS, Linux and a web app, each with a store listing, installer or browser link.
  • Local option: five catalogued on-device models, from Qwen3 0.6B to Gemma 4 12B, downloaded on demand.
  • Pace: four 3.x releases between 2 August and 2 October 2026, per the changelog.

📍 In One Sentence

Kai 9000 is a free, open-source AI assistant for phones, desktops and the browser that uses cloud providers by default and can run downloaded Gemma 4 models on the device.

💬 In Plain Terms

You install it from a store or installer, chat at once on a built-in free tier or add your own provider keys, and only switch to offline use if you download a model in settings.

📌Note: This review is based on the project's website, documentation, repository, privacy policy and store listings, checked on 30 October 2026. PromptQuorum has not installed, tested or benchmarked the app.

Read in:
🇺🇸en🇩🇪de🇫🇷fr🇯🇵ja🇨🇳zh🇪🇸es🇧🇷pt🇸🇦ar🇰🇷ko

Kai 9000 is a free, Apache-2.0 AI assistant from Berlin developer Simon Schubert that runs on Android, iOS, Windows, macOS, Linux and in the browser. It talks to cloud providers by default and can optionally run downloaded models on the device. This review sticks to the project's website, documentation, repository and store listings, checked on 30 October 2026; PromptQuorum has not run it.

Kai 9000 Review: Free Open-Source AI Assistant, Cloud or On-Device

What Is Kai 9000?

Kai 9000 is an AI assistant that remembers things between chats and can act through tools, not only reply with text. The documentation lists persistent memory, an editable system prompt the project calls a soul, a background heartbeat that checks in periodically, tools such as web search, calendar events, notifications and shell commands, MCP server support, and an interactive mode in which the model builds screens such as quizzes or dashboards.

It is published by Simon Schubert of Berlin under the Apache 2.0 licence on GitHub, as Kai 9000 on the website and as "Kai - AI" in the app stores. The repository describes it as an OpenClaw alternative for your pocket.

Get It

Kai 9000 is a normal download on every platform: store apps for phones, installers for desktop and a browser version, with no terminal step needed.

Item
Details
AndroidGoogle Play, F-Droid or APK on GitHub
iPhone and iPadApp Store, iOS 15.3 or later; Apple-silicon Macs too
WindowsMSI on GitHub Releases or winget
macOSDMG on GitHub Releases or Homebrew
LinuxDEB, RPM, AppImage, Flatpak or tarball on GitHub
Browserkai9000.com/app
PriceFree (Apache 2.0)
Websitekai9000.com
Documentationkai9000.com/docs
Privacy policyKai privacy policy
Source codeGitHub repository

Version reviewed: 3.3.0, released on 2 October 2026 per GitHub Releases and listed on F-Droid; the repository's latest code push was on 24 October 2026. The App Store listing showed 3.0.0 when checked, so iPhone users may be on an older build. Command-line installs from the README: brew install --cask simonschubert/tap/kai, winget install SimonSchubert.Kai, yay -S kai-bin.

On-Device Models and Limits

On-device use is optional and limited to a short catalogue of models that the app downloads from Hugging Face; the project's documentation, last verified 21 October 2026, lists them.

Model
Download
Note from the docs
Gemma 4 E2B IT2.59 GBTool calling reliable
Gemma 4 E4B IT3.66 GBTool calling reliable
Gemma 4 12B IT6.88 GBCannot load on iOS
LFM2.5 1.2B Instruct736 MBSmallest model with a tool template
Qwen3 0.6B614 MBChat only in practice

Limits named in the docs: text input only on the device, no interactive screens, a shortened tool list, a 64-bit Android device or, on Linux, a CPU with AVX2, and no local engine in the web build. The docs give no minimum RAM figure.

How to Get Started

The steps follow the project's getting-started page; PromptQuorum has not run them.

  1. 1
    Install the app
    Why it matters: Pick the store app or installer for your device from the Get It table.
  2. 2
    Chat on the Free tier
    Why it matters: Open the app and chat at once on the built-in Free tier, which needs no API key.
  3. 3
    Add a provider
    Why it matters: In Settings add a service such as OpenAI, Gemini or Anthropic, or an OpenAI-compatible address; several can be ordered into a fallback chain.
  4. 4
    Optionally go local
    Why it matters: In the Local Model card download a catalogue model or import a .litertlm file; only then does chat run on the device.
  5. 5
    Decide about tools
    Why it matters: Switch on tools such as shell, email or the Android Linux sandbox only if you want the assistant to act; the privacy policy says tools are off by default and need your permission.

Features Listed in the Docs

Every item below comes from the project's README and documentation.

  • Memory and personality. Persistent memory that the assistant stores and recalls, plus an editable system prompt.
  • Tools and MCP. Web search, notifications, calendar events, shell commands, email and scheduled tasks on remote models, plus connections to MCP servers.
  • Interactive screens. The model can build quizzes, dashboards and recipe cards that you navigate by tapping; remote models only.
  • Providers and fallback. Around 30 cloud providers and any OpenAI-compatible address, tried in the order you set.
  • Android Linux sandbox. An optional Alpine or Debian environment where the assistant can run commands and scripts without root.
  • Extras. Text to speech and image attachments to remote models.

Data Flow and Storage

What leaves the device depends on the model you pick. With a cloud provider or the built-in Free tier, your messages and the memories added to each prompt go to that service. The privacy policy says Kai has no account, no analytics and no servers of its own that receive your data, that the provider's terms apply, and it names Mistral AI's free tier as an example, saying that under that tier's terms inputs and outputs may be used for training. It adds that the location tool sends your IP address to ip-api.com and that tools reading SMS or calendar data pass that data to the AI service. With an on-device model, the docs say inference stays on the device.

On storage, the README and website say conversations are encrypted, while the project's encryption documentation says that on Android, iOS and desktop they sit in a local SQLite database protected by the operating system's app sandbox, with API keys and passwords in encrypted settings. F-Droid labels the app with the anti-feature Non-Free Network Services. The policy is dated 3 February 2026, before the 3.x releases, and the sources we read do not say who runs the Free tier's proxy or what it retains.

📌Note: These are the project's own statements; PromptQuorum has not audited the app's network traffic or storage.

Who Should Use It

  • People who want one assistant on every device with a free open-source licence and the choice of cloud, their own server or an on-device model.
  • Tinkerers who like memory, tools and MCP and will read the docs to switch features on.
  • Ollama or LM Studio users who want a phone or desktop front end for their own server, since a custom base address is supported.

What We Could Not Verify

  • Hands-on quality. PromptQuorum did not run the app, so answer quality, speed, battery and memory use of the on-device models are not assessed.
  • The Free tier. Its operator, data retention, usage limits and models are not described in the sources we read.
  • The iPhone build. The App Store showed 3.0.0 while other channels showed 3.3.0, so which features iPhone users get is unclear.
  • Independent checks. We found no third-party security or privacy audit of the app.

Competitors and Alternatives

Msty

Platforms:
Windows, macOS, Linux
Price / license:
Freemium / closed source
Key difference:
Closed-source desktop client for local and cloud models
Articles about Msty (8)

Also mentioned in:

+8 more not shown

OpenClaw

Platforms:
Windows, macOS, Linux
Price / license:
Free / MIT
Key difference:
Open-source agent framework; Kai's repository calls itself a pocket-sized alternative
Articles about OpenClaw (6)

Also mentioned in:

+3 more not shown

Competitor details change often; confirm each app's current price, license and platforms on its own page.

Compare in directory

Frequently Asked Questions

Can it use my own Ollama server?

Yes. The OpenAI-compatible service defaults to localhost:11434/v1, the usual Ollama address, and accepts any custom base address; on Android 17 or later the app asks for local-network permission before reaching a LAN host.

What happens when a provider fails?

Kai retries twice, then tries the next configured service and shows which one answered. An on-device failure is never silently handed to the cloud, and a local model is never used as a fallback target.

Is there a hosted web version?

Yes, at kai9000.com/app; because the web build has no local engine, it needs a remote provider or the Free tier.

Verdict

Kai 9000 is a free, Apache-2.0 assistant for phones, desktops and the browser that treats memory and tools as core features and lets you choose between cloud providers, your own server and downloaded on-device models. The trade-offs are that the local path is optional and narrower than the cloud path, the project's own pages differ on how conversations are stored, and PromptQuorum could not test it. It suits people who want one open-source assistant across devices; readers who want a phone app built around local models can compare PocketPal AI, and desktop users Jan.

Sources

← Back to Power Local LLM

MCP