Turn natural language into CLI commands using Apple's on-device AI, powered by OpenCode.

LocalCode is a proof-of-concept that integrates Apple Foundation Models (AFM) with OpenCode as a local AI provider. Tell it what you want in plain English, it suggests the right command via tool calls, and you approve before execution.

All AI processing happens locally on your Mac. No cloud, no data leaving your machine.

• Apple Silicon Mac (M1/M2/M3/M4)
• macOS 26+
• Xcode 26+ (for building the Swift helper)
• Bun 1.3+

LocalCode/
├── LocalCode/Sources/afmhelper/   # Swift AFM helper
│   └── main.swift                # Apple FoundationModels integration
├── start-afm-server.sh           # HTTP middleware (Bun)
└── pre-commit.sh                 # Pre-commit hook

No fork needed - uses global OpenCode with provider config in ~/.config/opencode/opencode.json

Flow:

1. AFM Server wraps Swift helper with OpenAI-compatible API
2. OpenCode uses AFM as a provider via @ai-sdk/openai-compatible
3. AFM returns command suggestions as tool calls
4. OpenCode shows command approval UI → user approves → command executes

# 1. Start the AFM middleware server
./start-afm-server.sh &

# 2. Configure OpenCode provider in ~/.config/opencode/opencode.json:
{
  "provider": {
    "localcode-afm": {
      "npm": "@ai-sdk/openai-compatible",
      "name": "LocalCode AFM",
      "options": {
        "baseURL": "http://localhost:8080/v1"
      },
      "models": {
        "afm": { "name": "Apple Foundation Models" }
      }
    }
  }
}

# 3. Run OpenCode
opencode

# 4. Select "LocalCode AFM" provider via /models command

You: "list all python files in this directory"
OpenCode (AFM): [Tool Call: bash { command: "find . -name '*.py'" }]
                 ↑ Approval UI appears with approve/reject buttons
User: clicks approve
OpenCode: executes find . -name '*.py'
Output: ./file1.py
        ./subdir/file2.py

AFM acts as a command translator - it takes natural language and produces shell commands as tool calls that OpenCode can display with approval UI.

File Operations:

• "list all python files"
• "find all files named hello"
• "show me the largest files in this directory"
• "count all files in this directory"

System & Network:

• "check if port 8080 is in use"
• "show git status"

Search:

• "grep for hello in this directory"

# Check server is running
curl http://localhost:8080/v1/models

# Test chat completion (returns tool call)
curl -X POST http://localhost:8080/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{"model":"afm","messages":[{"role":"user","content":"hello"}]}'

# Test streaming
curl -X POST http://localhost:8080/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{"model":"afm","messages":[{"role":"user","content":"hello"}],"stream":true}'

Working:

• ✅ AFM server with OpenAI-compatible API
• ✅ Tool call responses for command approval UI
• ✅ SSE streaming support
• ✅ OpenCode provider integration via @ai-sdk/openai-compatible
• ✅ Extracts user message from OpenCode's system prompt

QA Failing:

• ❌ OpenCode streaming shows text streaming down instead of command approval dialog
• ❌ AFM outputs partial commands when receiving OpenCode's full system prompt
• ❌ Further debugging needed on OpenCode integration

Run pre-commit hook manually:

./pre-commit.sh

Or install as git hook:

cp pre-commit.sh .git/hooks/pre-commit && chmod +x .git/hooks/pre-commit
git config core.hooksPath .git/hooks

Apple's Foundation Models framework is new and under-documented. This project proves it's viable for CLI tools and provides a reference implementation for others building on AFM with OpenCode.

POCs have rough edges. Contributions welcome:

1. Fork and create a feature branch
2. Make your changes
3. Open a PR

MIT