Sidekick
A native macOS app for private, offline chat with local LLMs that can reference your files, folders, and websites.
Evidence shows local-first and offline data processing, but permission details are not explicit in source. User confirmation mechanisms exist (e.g., expert activation, drag-and-drop files) but not detailed. Data flow transparency limited; no privacy policy or data flow diagram. Sensitive data handling (contacts, API keys) not clearly specified. Dependency security lacks dependency list or vulnerability scan. External effects (e.g., sending emails, generating files) exist but user confirmation not detailed. Rollback not mentioned. Source attribution partially via README credits.
Self-consistency: README features align with screenshots, but no test evidence. Dependency availability: relies on llama.cpp etc., but no version pinning or mirrors. Failure messages: no documentation of error handling or user prompts.
Target audience clear (students, researchers), scenarios diverse (research, writing, coding). Capability boundaries partially described (local models, APIs), but limitations not explicit. Trigger precision: function calling and image generation auto-detected, but false trigger risk not discussed. Environment fit: only Apple Silicon, no other platforms.
Information architecture clear, with documentation links. Install notes provide Homebrew and manual steps. Naming stability: version numbers not mentioned. Examples and FAQ: examples present but no FAQ. Known limitations: not explicitly listed. License: MIT, but copyright holder unverified. Versioning/changelog missing. Maintenance responsibility: contact provided, but no maintenance plan.
Output usability: provides citations, code export, etc., but not verified. Marginal value: local-first, RAG, function calling offer unique value. Cost-benefit: free and open source, but requires Apple Silicon and RAM, no performance data.
Claim traceability: README has screenshots and links, but no test results. Cross-source corroboration: no third-party verification. Fact-inference separation: not explicitly distinguished.
- No dependency list or vulnerability scan provided; dependency security unknown.
- Security measures for sensitive data (e.g., contacts, API keys) not explicit.
- User confirmation for external actions (e.g., sending emails, generating files) not detailed.
- Versioning/changelog missing, making updates hard to track.
- Only supports Apple Silicon, limiting environment fit.
What does this agent do, and when should you use it?
Sidekick is a native macOS application that lets you chat with a local LLM powered by the built-in llama.cpp engine, enabling responses grounded in your local files, folders, and websites without installing any additional software. It supports modern GGUF models like Qwen3.5 out of the box, and also accommodates OpenAI-compatible APIs for remote model options. Sidekick features RAG-based 'experts' for resource management, function calling, a Deep Research agent, memory, a canvas for editing code/websites, and image generation (requires macOS 15.2+ and Apple Intelligence). All conversations default to offline, keeping your data secure. Installation is via Homebrew or direct download, and the app is designed for Apple Silicon Macs with at least 8GB RAM.
Sidekick loads and runs local GGUF models (e.g., Qwen3.5) through its llama.cpp backend, and can also connect to remote models via OpenAI-compatible APIs, with presets for providers like OpenAI, Anthropic, Google AI Studio, DeepSeek, Groq, and more (using your own API key). Users create 'experts' and attach resources (folders, files, URLs) to each; Sidekick uses RAG to retrieve relevant information from activated experts when answering queries. It also supports temporary context by dragging files into the input box, and can perform web searches to retrieve up-to-date information. Function calling allows sequential tool invocations (e.g., calculating metrics, drafting emails) until a result is achieved, enabling fully local agent workflows. The built-in Deep Research agent can read 50-80 web pages and synthesize a research report. Sidekick features persistent memory across conversations, a Canvas for editing and previewing websites/code, and automatic image generation based on text prompts (via CoreML). It renders Markdown with LaTeX, data visualizations (bar, line, pie charts), and syntax-highlighted code.
- A student can create experts for different subjects (English Literature, Mathematics, Computer Science), activate the relevant one, and ask questions like 'Did the Aztecs use captured Spanish weapons?' — Sidekick returns direct quotes with page numbers and opens the referenced paper on click.
- A researcher can use the Deep Research agent to specify a research topic and let Sidekick autonomously read dozens of web pages and produce a synthesized report.
- A professional can leverage function calling to ask Sidekick to calculate financial metrics (e.g., Nvidia Q3 2025) — it makes 27 tool calls, saves a CSV file, and presents the results.
- A privacy-conscious user can process sensitive local documents entirely offline, without uploading data to the cloud.
- A writer can use Canvas to create, edit, and preview websites or code, with the ability to select text and prompt selective edits.
- A casual user can drag a file into the input to ask questions about its contents, with no setup required.
What are this agent's strengths and limitations?
- Truly local-first: runs offline, keeping your data on your device by default.
- All-in-one feature set: RAG, function calling, deep research, memory, canvas, and image generation.
- Supports modern GGUF models and presets for major cloud providers, offering flexibility.
- Built-in llama.cpp backend optimized for Apple Silicon, delivering fast generation.
- Drag-and-drop file access makes it easy to add context without configuration.
- Only works on Apple Silicon Macs; no Intel or non-macOS support.
- Image generation requires macOS 15.2+ and Apple Intelligence, excluding older systems.
- Advanced features like function calling and deep research may be resource-intensive on lower-RAM devices.
- While it supports remote APIs, the core experience is local-model-centric, and remote integration may be less polished than dedicated apps.
- The project is under active development, so stability and documentation may lag.
How do you install or deploy this agent?
- Ensure you have an Apple Silicon Mac (RAM ≥ 8GB).
- Install via Homebrew:
brew install --cask arcadi4/tap/sidekick. - Alternatively, download the latest release from the GitHub Releases page and move it to Applications.
- For developers: clone the repository, run
./setup.sh <TEAM_NAME>to set the Xcode project team, then open and run in Xcode.
How do you use this agent?
- Launch Sidekick. On first run, you'll be prompted to select a local model (GGUF format) or configure an API key for remote models.
- Create 'experts' for different topics and add relevant folders or websites; activate the expert(s) you want Sidekick to use.
- Type your question in the input field. Sidekick will retrieve context from the activated experts using RAG and generate an answer with citations.
- Enable Web Search if you need the latest information.
- For deep research, start the Deep Research feature and enter your research topic; Sidekick will work through multi-step research and produce a report.
- Drag files into the input field to add them as temporary resources.
- To generate images, describe the image you want; Sidekick automatically detects the need and generates it (requires macOS 15.2+).