A self-contained, private AI assistant for your Mac. Sumika runs local models directly through MLX — no Ollama, LM Studio, external model server, API key, or separate runtime required.
Install the app, choose and download a model, and start chatting right away.
Use Chat mode to write, translate, summarize, research, or code. Switch to Agent mode when you want Sumika to write code, work with your files, use approved tools, or connect to local apps. Your conversations and model execution stay on your Mac, and you remain in control of every action.
- 💬 Everyday AI assistance: write, brainstorm, translate, summarize, research, and ask questions without needing a technical background.
- 🏠 Everything you need in one app: download a model from the built-in model browser and start chatting. No Ollama, LM Studio, or separate inference backend is required.
- 🌟 No recurring subscription: use local models without paying for a hosted AI assistant plan.
- 🧭 Explicit context: attach files, focus workspace context, and inspect what the model sees before the workflow grows opaque.
- 🛠 An agent that asks first: let Sumika work with files, use connected tools, and run commands while you review sensitive actions.
- 🧩 Connect local apps: extend Agent mode through Model Context Protocol (MCP) servers and choose the integrations available to each session.
- ✅ Review before action: writes, edits, shell commands, and web access pass through approval instead of running as hidden automation.
- 🌐 Bring your own search: connect a self-hosted SearXNG instance or use the built-in DuckDuckGo search provider.
- 📄 Bring your own fetcher: keep the built-in page extractor or point fetch at a self-hosted Firecrawl instance.
- 🧰 Terminal and browser built in: run approved workspace commands and inspect local previews without leaving the app.
- 🖥 Build and preview locally: create small apps, prototypes, and HTML experiments, then inspect them beside the chat.
- 🗣 Speak and dictate: listen to assistant responses with Apple system voices and turn speech into prompts with local English or multilingual transcription models.
- 🔄 Update checks built in: automatically check for signed releases and install them through the native Sparkle update flow.
- 🧾 Inspectable transcript: keep prompts, assistant responses, tool calls, approvals, and command output visible in the chat.
- Download the latest
Sumika-*-macos.dmgfrom the GitHub Releases page. - Open the downloaded DMG and drag Sumika into Applications.
- Start Sumika from the Applications folder and download a local model from the Models screen.
- Open a conversation and choose Chat or Agent mode.
Release downloads are signed and notarized for macOS. Once installed, Sumika checks automatically for new versions. When an update is available, the app guides you through installing it, so you do not need to download another DMG. You can also check manually from Sumika > Check for Updates….
All listed models run locally and support Chat mode, Agent tool calling, and image input.
| Model | Download size |
|---|---|
| Gemma 4 E4B QAT 4-bit | 6.8 GB |
| Gemma 4 12B QAT 4-bit | 11.0 GB |
| Gemma 4 26B QAT 4-bit | 15.6 GB |
| Gemma 4 31B QAT 4-bit | 28.8 GB |
| Qwen 3.6 27B 4-bit | 16.1 GB |
| Qwen 3.6 27B 8-bit | 29.5 GB |
| Qwen 3.6 35B A3B 4-bit | 20.4 GB |
| Qwen 3.6 35B A3B 8-bit | 37.7 GB |
Sumika can work in agent mode, write files through approval-aware tools, and keep the transcript inspectable while it works.
Build small local HTML, CSS, and JavaScript prototypes, then inspect them in the native preview pane.
Download, load, and inspect local models from the macOS app without turning the chat into a cloud workflow.
- Write and refine text, brainstorm ideas, translate languages, and summarize longer content.
- Ask questions about your own files and choose what the model can see.
- Research topics on the public web through reviewable search and fetch tools.
- Connect local apps and services through configured MCP integrations when you want Sumika to take action beyond the chat.
- Let the agent read, organize, search, and update local project files while you review sensitive actions.
- Build small apps, scripts, games, and UI prototypes in short, reviewable steps.
- Review generated file writes, file edits, shell commands, and workspace diffs before they run.
- Use the integrated terminal and browser preview while working through an agent task.
- Open local HTML previews and inspect browser state while iterating.
- Transcribe speech locally and dictate prompts in English, German, and other supported European languages instead of typing them.
- Listen to assistant responses with installed Apple voices.
- Receive automatic update checks and install signed releases from the app.
- Follow prompts, assistant responses, tool calls, approvals, and command output in one visible transcript.
Choose how much Sumika can do in each conversation:
- Chat: talk, write, translate, summarize, and research the public web. Chat mode cannot access local files, run commands, or make changes.
- Agent: let Sumika work with a selected folder, connected tools, commands, and browser previews. Sensitive actions remain subject to approval.
You select the mode yourself. Sumika never grants itself access because of how a prompt is worded.
Sumika is designed as a private alternative to subscription-based cloud assistants.
- No recurring AI subscription or hosted workspace account is required.
- No telemetry, prompts, transcripts, commands, or workspace contents are exported by the app.
- Model execution, chat history, speech output, and dictation stay on your Mac.
- Network access is explicit: web search and fetch tools only run when available in the selected mode and approved by policy.
- You can use the built-in DuckDuckGo search provider or point Sumika at your own SearXNG instance. Fetch uses the built-in extractor by default and can optionally use a self-hosted Firecrawl instance without storing an API key.
Sumika includes two local voice surfaces:
- Assistant speech adds play controls to completed text responses. It uses Apple system voices installed on the Mac, supports language and voice selection, and lets you tune speech rate.
- Speech-to-text and composer dictation record and transcribe prompts on the Mac. Choose the small, fast English model for English prompts or the larger multilingual Parakeet model to capture prompts in German and other supported European languages.
The Model Context Protocol (MCP) lets AI assistants use tools provided by other apps and services. Configure MCP servers globally in Settings using stdio or Streamable HTTP, then choose the servers available to each Agent session from the composer. MCP tools stay out of Chat mode, and every MCP tool call requires approval before execution.
Most AI assistants send conversations to cloud models and charge a recurring subscription. Sumika takes a different approach:
- Local-first model execution on macOS
- Everyday assistance without requiring technical knowledge
- No recurring AI subscription
- User-controlled workspace context
- Reviewable agent steps instead of hidden automation
- Approval-gated tool and shell execution
- Visible transcripts and tool states for review
- Native macOS workflows instead of a browser-first interface
sumika means "dwelling" or "place to live" in Japanese. sumika.chat is meant
as a local home for AI agents: close to your files, explicit about what context
they see, and reviewable before they act.
Sumika is an evolving prototype. It is useful for local conversations, everyday AI assistance, and agent workflows, but APIs, persisted data, and workflows are still changing.
The project uses one Swift package with three production modules and a thin native Xcode app target:
SumikaCoreowns provider-neutral domain models, agent/workflow logic, persistence, tools, and runtime interfaces.SumikaRuntimeMLXimplements the local MLX/Hugging Face runtime behind the core interfaces.SumikaAppowns SwiftUI/AppKit, launch composition, and macOS integrations such as Sparkle.sumika/contains only the native app launcher, resources, entitlements, and bundle metadata.
Dependencies point one-way: SumikaApp -> SumikaRuntimeMLX ->
SumikaCore; SumikaApp may also use SumikaCore directly. All external
Swift package dependencies are declared in the root Package.swift; the Xcode
app target links only the local SumikaApp product.
- Tool Runtime: core flow for adding type-safe tools, permissions, registries, and model-facing tool calls.
- Chat Runtime: chat turn lifecycle, cancellation, transcript state, and model-context filtering.
Install the local task runner, linter, and formatter:
brew install just swiftlint swift-formatBuild the app locally:
just build
open "build/DerivedData/Build/Products/Debug/Sumika.app"The first Xcode build may ask you to enable the pinned MLXHuggingFaceMacros
package macro. Review and approve it in Xcode. Hosted CI cannot approve package
plugins or macros interactively and uses the explicit validation-skip flags in
the project task runner.
Build an unsigned release app:
just release-unsigned
open "build/DerivedData/Build/Products/Release/Sumika.app"Build and export a Developer ID-signed release archive:
DEVELOPER_ID_APPLICATION="Developer ID Application: …" just release-signedThe signed release command verifies the exported app, including the embedded Sparkle framework, updater, XPC helpers, update feed configuration, nested signatures, signing team, hardened runtime, and release entitlements.
You can also build from Xcode by opening Sumika.xcodeproj and running the
Sumika scheme for macOS.
Common development tasks:
just test
just lint
just format
just final-checkjust build and just release-unsigned run the Sumika Xcode scheme with a
stable DerivedData path under build/DerivedData. just test runs every unit
and integration test target through SwiftPM; Xcode remains responsible for the
app launcher/resources and UI tests. just lint runs SwiftLint using
.swiftlint.yml. just format checks Swift sources with swift-format.
just final-check runs the broader local verification suite before review.
just resolve-packages resolves both the root SwiftPM graph and the Xcode app
graph, then synchronizes the Xcode pin states with the root resolution. Commit
both Package.resolved files after dependency changes. The root lockfile is the
canonical pin selection; the Xcode lockfile retains its workspace-specific
metadata.
just check-package-locks disables automatic dependency updates and verifies
that both committed lockfiles still satisfy their graph and resolve identical
package pins. The files represent different resolver roots and are not expected
to be byte-identical; metadata such as originHash and normalized repository
URLs may differ.
If a Dependabot PR changes the root graph and this check reports a stale Xcode
lockfile, run just resolve-packages and commit the regenerated Xcode
Package.resolved file to the PR.
Licensed under the Apache License 2.0.