A local AI coding assistant for macOS, powered by Apple MLX. No cloud. No subscriptions. No data leaving your machine.
MLX Code runs language models directly on Apple Silicon using the mlx-swift framework. It provides a chat interface with 14 built-in tools that can read files, search code, run shell commands, build Xcode projects, manage Git repos, and interact with GitHub -- all driven by a local model running on your GPU.
+-----------------------------------------------------------------------------------+
| MLX Code (macOS App) |
| |
| +---------------------------+ +--------------------------------------------+ |
| | SwiftUI Views | | ViewModels | |
| | | | | |
| | ChatView | | ChatViewModel | |
| | MessageRowView | | Conversation state & tool execution | |
| | CodeBlockView | | Streaming token display | |
| | CollapsibleToolResult | | Slash command dispatch | |
| | ToolApprovalView | | | |
| | ThinkingIndicatorView | | ProjectViewModel | |
| | CommandPalette | | Build, archive, deploy operations | |
| | PromptTemplatesView | | | |
| | SettingsView | | GitHubViewModel | |
| | OnboardingView | | Issues, PRs, branches | |
| | DiffView | | | |
| | GitHubPanelView | | CodeAnalysisViewModel | |
| | CodeAnalysisPanelView | | Metrics, dependencies, lint | |
| +---------------------------+ +--------------------------------------------+ |
| | | |
| | User messages | |
| +------------------+-------------------+ |
| | |
| v |
| +-------------------------------------------------------------------------+ |
| | Tool Execution Layer | |
| | | |
| | ToolRegistry (14 tools, 2 tiers) | |
| | +-----------+ +-----------+ +----------+ +----------+ +---------+ | |
| | | File Ops | | Bash | | Grep | | Glob | | Edit | | |
| | +-----------+ +-----------+ +----------+ +----------+ +---------+ | |
| | CORE CORE CORE CORE CORE | |
| | | |
| | +-----------+ +-----------+ +----------+ +----------+ +---------+ | |
| | | Xcode | | Git | | GitHub | | CodeNav | |CodeAnal.| | |
| | +-----------+ +-----------+ +----------+ +----------+ +---------+ | |
| | +-----------+ +-----------+ +----------+ +----------+ | |
| | | ErrorDiag | | TestGen | |DiffPreview| | Help | | |
| | +-----------+ +-----------+ +----------+ +----------+ | |
| | DEVELOPMENT tier (available when a project is open) | |
| +-------------------------------------------------------------------------+ |
| | |
| v |
| +-------------------------------------------------------------------------+ |
| | Inference Engine | |
| | | |
| | MLXService (Swift actor) | |
| | - Native mlx-swift-lm: LLMModelFactory + ModelContainer | |
| | - Chat templates via tokenizer (Jinja), flat-prompt fallback | |
| | - AsyncStream<Generation> for token streaming | |
| | - SafeTensors-only model loading (PyTorch pickle rejected) | |
| | - Native HuggingFace Hub downloads (no Python) | |
| | | |
| | ContextManager (actor) ContextBudget | |
| | - Token estimation - Budget allocation: | |
| | - Message compaction 70% messages / 20% project / | |
| | - Project context inclusion 10% summary | |
| | | |
| | SystemPrompts UserMemories (actor) | |
| | - ~500 token tool prompt - 50+ built-in coding standards | |
| | - Compact tool descriptions - 8 categories of rules | |
| | - Runtime memory injection - Custom memories stored locally | |
| +-------------------------------------------------------------------------+ |
| |
| +-------------------------------------------------------------------------+ |
| | Security & Utilities | |
| | | |
| | CommandValidator ModelSecurityValidator KeychainManager | |
| | - Regex pattern block - SafeTensors only - Secure storage | |
| | - Shell metachar filter - Trusted sources - API key mgmt | |
| | - Audit logging - Hash verification | |
| | | |
| | SecureLogger SecurityUtils RepetitionDetector | |
| | - Category-based logs - Input sanitization - Loop detection | |
| | - No PII in logs - Command validation - Auto-break | |
| +-------------------------------------------------------------------------+ |
| |
+----------|---------------------|-----------------------|--------------------------+
| | |
v v v
+-------------------+ +-------------------+ +-------------------------+
| Xcode Extension | | Desktop Widget | | Nova API Server |
| | | | | |
| 5 Editor commands | | Small / Med / Lrg | | HTTP on 127.0.0.1:37422 |
| Shared App Group | | Model status | | |
| mlxcode:// scheme | | Token speed | | /api/status |
| | | Memory usage | | /api/chat |
| Explain Selection | | Quick actions | | /api/conversations |
| Refactor | | | | /api/model |
| Generate Tests | | WidgetKit | | /api/metrics |
| Fix Issues | | 5-min refresh | | /api/prompts |
| Ask MLX Code | | | | /api/cancel |
+-------------------+ +-------------------+ +-------------------------+
Key design decisions:
- Pure Swift inference --
mlx-swift-lmframework, no Python dependency anywhere in the pipeline - Actor isolation --
MLXService,ContextManager, andUserMemoriesare Swift actors for thread safety - SafeTensors only -- PyTorch pickle files (.bin, .pt) are rejected at both discovery and load time
- Compact prompts -- Tool descriptions fit in ~500 tokens, not 4,000, leaving room for conversation
- Two tool tiers -- Core tools always available; development tools appear when a project is open
- Budget-aware context -- Token allocations: 70% recent messages, 20% project context, 10% summary
Type a message, and the model reads files, searches code, runs commands, and builds your project. The 14 built-in tools cover the full development workflow:
| Tool | Purpose | Tier |
|---|---|---|
| File Operations | Read, write, edit, list, delete files | Core |
| Bash | Run shell commands | Core |
| Grep | Search file contents with regex | Core |
| Glob | Find files by pattern | Core |
| Edit | Apply targeted file edits | Core |
| Xcode | Build, test, clean, archive, full deploy pipeline | Dev |
| Git | Status, diff, commit, branch, log, push, pull | Dev |
| GitHub | Issues, PRs, branches, credential scanning | Dev |
| Code Navigation | Jump to definitions, find symbols | Dev |
| Code Analysis | Metrics, dependencies, lint, symbols | Dev |
| Error Diagnosis | Analyze and explain build errors | Dev |
| Test Generation | Create unit tests from source files | Dev |
| Diff Preview | Show before/after file changes | Dev |
| Help | List available commands and usage | Dev |
Read-only tools (grep, glob, file read, code navigation) auto-approve. Write and execute tools (bash, file write, xcode build) require confirmation.
/commit /review /test /docs
/refactor /explain /optimize /fix
/search /plan /help /clear
Select code in Xcode and invoke from Editor > MLX Code:
- Explain Selection -- understand what code does
- Refactor Selection -- get an improved version
- Generate Tests -- write unit tests for the selection
- Fix Issues -- find and fix bugs
- Ask MLX Code -- open with code pre-loaded, ask anything
The extension communicates via a shared App Group container and the mlxcode:// URL scheme.
15 curated prompt templates across 9 categories, engineered for local LLMs (Qwen, Phi, Mistral). Browse templates, fill in variables, preview the rendered prompt, and send -- all without leaving the app.
| Category | Templates |
|---|---|
| Review | Swift Code Review, Review PR / Diff |
| Debug | Fix a Bug, Fix Build Error |
| Generate | Implement Feature, Add API Endpoint, Create Data Model |
| Refactor | Refactor File |
| Test | Add Unit Tests |
| Document | Add Documentation, Explain Code |
| Security | Security Audit, Add Error Handling |
| Performance | Performance Analysis |
| Deploy | Migration Plan |
Three sizes (small, medium, large) showing model status, token speed, memory usage, and quick-action buttons that deep-link into the app.
The GitHub tool connects to the GitHub API to view and create issues, list and create pull requests, manage branches, and scan for exposed credentials before pushing.
Persistent preferences that shape assistant behavior across conversations. 50+ built-in coding standards across 8 categories (personality, code quality, security, Xcode, git, testing, docs, deployment), plus custom memories stored locally. Injected into the system prompt at runtime -- never hardcoded.
- Token budgeting with automatic message compaction when the context fills up
- Real-time context window usage bar synced to the actual model's context size
- Project context auto-included when a workspace is open
- Budget allocation: 70% recent messages, 20% project context, 10% summary of older messages
Code blocks render with syntax highlighting for Swift, Python, JavaScript, TypeScript, Bash, JSON, and Objective-C. Keywords, types, strings, comments, and numbers are all colored.
MLX Code exposes a local HTTP API on port 37422 for integration with external tools and automations. The server binds to 127.0.0.1 only -- no external network exposure.
GET /api/status App status, model loaded state, uptime
GET /api/ping Health check
GET /api/conversations List all conversations
GET /api/conversations/:id Single conversation with messages
POST /api/conversations Create new conversation
DELETE /api/conversations/:id Delete conversation
POST /api/chat Send a message, get response {"message": "..."}
GET /api/model Current model info
POST /api/model/load Load a model {"model": "path"}
GET /api/metrics Performance metrics (tokens/sec, memory)
POST /api/cancel Cancel current generation
GET /api/prompts List all prompt templates
GET /api/prompts/:id Get a template with variables
POST /api/prompts/render Render template with filled variables
# Check status
curl -s http://127.0.0.1:37422/api/status | python3 -m json.tool
# Send a message
curl -X POST http://127.0.0.1:37422/api/chat \
-H "Content-Type: application/json" \
-d '{"message": "List all Swift files in this project"}'
# Browse prompt templates
curl -s http://127.0.0.1:37422/api/prompts | python3 -m json.toolThe API starts automatically when the app launches.
MLX Code uses mlx-community models from Hugging Face, quantized for Apple Silicon.
| Model | Size | Context | Best for |
|---|---|---|---|
| Qwen 2.5 7B (default) | ~4 GB | 32K | General coding, tool calling |
| Mistral 7B v0.3 | ~4 GB | 32K | Versatile, good at instructions |
| DeepSeek Coder 6.7B | ~4 GB | 16K | Code-specific tasks |
| Qwen 2.5 14B | ~8 GB | 32K | Best quality (needs 16GB+ RAM) |
Models download automatically on first use via the native Hub Swift API. You can also add custom models from any mlx-community repository.
All models must be in SafeTensors format. PyTorch pickle files (.bin, .pt) are rejected at both discovery and load time.
- macOS 14.0 (Sonoma) or later
- Apple Silicon (M1, M2, M3, M4)
- 8 GB RAM minimum (16 GB recommended for 7B models)
- No Python required -- inference and downloads are pure Swift
- Xcode 15+ -- only needed for the Source Editor Extension feature
- Download
MLXCode-vX.Y.Z.dmgfrom Releases - Open the DMG and drag MLX Code to
/Applications - Launch the app
- Go to Settings > Models and download a model
- Load the model and start chatting
- Open System Settings > Privacy & Security > Extensions > Xcode Source Editor
- Enable MLX Code
- Restart Xcode
- Select code, then use Editor > MLX Code menu
See INSTALLATION.md for the full setup guide, including model paths, troubleshooting, and building from source.
git clone git@github.com:kochj23/MLXCode.git
cd MLXCode
open "MLX Code.xcodeproj"Build the MLX Code scheme in Xcode 15+ targeting macOS 14.0+. The project uses Swift Package Manager for dependencies (mlx-swift, mlx-swift-lm, Hub).
| Dependency | Purpose |
|---|---|
| mlx-swift | Apple MLX framework bindings for Swift |
| mlx-swift-lm | LLM inference: LLMModelFactory, ModelContainer, GenerateParameters |
| MLXLMCommon | Shared types: UserInput, LMInput, AsyncStream<Generation> |
| Hub | HuggingFace model downloads via HubApi.snapshot() |
| SwiftUI | All views, settings, onboarding, panels |
| WidgetKit | Desktop widget (3 sizes) |
| XcodeKit | Source Editor Extension |
| Network (NWListener) | Nova API HTTP server on loopback |
| CryptoKit | Model file hash verification |
| Security | Keychain storage for API keys |
MLXServiceis a Swift actor -- all model state (load, unload, generate) is automatically serializedContextManagerandUserMemoriesare also actors- Token streaming uses
AsyncStream<Generation>, delivered to@MainActorviaMainActor.run - All background loops check
Task.isCancelledfor clean shutdown - Inference breaks immediately when
</tool>is detected in output, preventing the model from running to maxTokens
Shell Execution Safety:
CommandValidatorblocks dangerous patterns (rm -rf /, fork bombs,sudo,eval,curl | sh) using regex word-boundary matching- Shell metacharacters (
;,|,&,$, backticks) are rejected before pattern matching - Git and build tools use
Process.currentDirectoryURLinstead of string interpolation - Write/execute tools require user confirmation; read-only tools auto-approve
Model Security:
- Only SafeTensors (.safetensors) format is permitted
- PyTorch pickle files (.bin, .pt) are blocked -- they can execute arbitrary Python code
ModelSecurityValidatorverifies trusted sources and file hashes via CryptoKit
Data Privacy:
- 100% local inference -- no prompts or responses leave your machine
- No telemetry, analytics, or crash reporting
- No cloud AI services (OpenAI, Anthropic, etc.)
- The only external network calls are to the GitHub API, explicitly invoked by the user
- User memories stored locally, never transmitted
Being honest about limitations:
- No web browsing -- cannot fetch URLs or browse the internet
- No image/video/audio generation -- this is a code assistant
- Small model constraints -- 3-14B parameter models make mistakes, especially with complex multi-step reasoning
- Tool calling is imperfect -- local models sometimes format tool calls incorrectly (JSON auto-repair and retry help, but are not perfect)
- Extension requires app switch -- the Xcode extension opens MLX Code in a separate window rather than responding inline
| Version | Date | Highlights |
|---|---|---|
| v6.4.0 | April 2026 | Prompt Engineering Toolkit -- 15 templates, 9 categories, API endpoints, in-app browser |
| v6.3.0 | March 2026 | Xcode Source Editor Extension, native downloads (no Python), syntax highlighting, collapsed tool calls, accurate context bar, resume generation, SafeTensors-only loading |
| v6.2.0 | March 2026 | Replaced Python subprocess with native mlx-swift-lm inference, removed 2,726 lines of dead code |
| v6.1.x | March 2026 | Security audit (31 findings resolved), Keychain migration, SecureLogger |
| v6.0.0 | February 2026 | GitHub integration, code analysis, Xcode deploy pipeline, user memories, 14 tools |
| v5.0.0 | February 2026 | Major simplification -- deleted 41 files (~16,000 lines), compact system prompt, Qwen 2.5 default |
| v4.0.0 | February 2026 | Chat templates, tool tier system, context budgeting, tool approval flow |
| v1.x | Jan-Feb 2026 | Initial release, MLX backend, desktop widget, basic chat |
- Deeper Xcode integration -- write responses back into the editor buffer without switching apps
- Structured output -- grammar-constrained generation to guarantee well-formed tool calls from smaller models
- Streaming download progress -- real-time progress bar for model downloads
MLX Code/
MLXCodeApp.swift App entry point
Models/ Data models (Message, Conversation, MLXModel, AppSettings)
Services/
MLXService.swift Native MLX inference engine (actor)
ContextManager.swift Token budgeting and message compaction
ContextBudget.swift Budget allocation ratios
NovaAPIServer.swift HTTP API on port 37422
GitHubService.swift GitHub API client
GitService.swift Local git operations
XcodeService.swift Xcode build/test/archive
UserMemories.swift Persistent user preferences
ConversationManager.swift Conversation persistence
FileService.swift File system operations
SessionManager.swift Session lifecycle
SlashCommandHandler.swift Slash command dispatch
XcodeActionHandler.swift Handles requests from Xcode extension
ModelSecurityValidator.swift SafeTensors validation + hash checks
Tools/
ToolRegistry.swift Central tool registry and dispatcher
ToolProtocol.swift Tool interface definition
ToolTier.swift Core vs. development tier classification
SystemPrompts.swift System prompt generation with tool + memory injection
PromptTemplates.swift 15 curated prompt templates
BashTool.swift Shell command execution
FileOperationsTool.swift File CRUD
GrepTool.swift Content search
GlobTool.swift File pattern matching
EditTool.swift Targeted file edits
XcodeTool.swift Build, test, archive, deploy
GitIntegrationTool.swift Git operations
GitHubTool.swift GitHub API tool
CodeNavigationTool.swift Symbol lookup and navigation
CodeAnalysisTool.swift Metrics, deps, lint
ErrorDiagnosisTool.swift Build error analysis
TestGenerationTool.swift Unit test generation
DiffPreviewTool.swift Before/after diff display
HelpTool.swift Command help
MemorySystem.swift Tool execution memory
Views/ SwiftUI views (ChatView, SettingsView, etc.)
ViewModels/ ChatViewModel, ProjectViewModel, GitHubViewModel, etc.
Utilities/
CommandValidator.swift Shell command security validation
SecureLogger.swift Category-based secure logging
KeychainManager.swift macOS Keychain wrapper
SecurityUtils.swift Input sanitization
RepetitionDetector.swift Detects and breaks inference loops
BuildErrorParser.swift Xcode build error parsing
MLX Code Extension/ Xcode Source Editor Extension (5 commands)
SourceEditorExtension.swift Extension entry point
SourceEditorCommand.swift Command handler, App Group IPC
MLX Code Widget/ WidgetKit desktop widget (3 sizes)
MLXCodeWidget.swift Small, medium, large widget views
SharedDataManager.swift App Group data sync
WidgetData.swift Widget data model
MLX Code Tests/ Unit tests
MIT License -- Copyright 2025-2026 Jordan Koch
See LICENSE for the full text.
This is a personal project created on my own time. It is not affiliated with, endorsed by, or representative of my employer.
Written by Jordan Koch