@appifex/mcp-server

@appifex/mcp-server

MCP server that exposes the full DTC design-to-code toolkit to AI agents over the Model Context Protocol.

Quick Start

# Build the server
pnpm build

# Run it (stdio transport — meant to be spawned by an AI agent)
node dist/index.js

Connecting from Pydantic AI (Python)

from pydantic_ai import Agent
from pydantic_ai.mcp import MCPServerStdio

dtc_server = MCPServerStdio(
    "node",
    args=["packages/appifex-dtc/packages/mcp-server/dist/index.js"],
    env={"HOME": os.environ["HOME"]},  # needed for ~/.dtc config
    timeout=30,
    read_timeout=600,  # fix loop and pipeline can take minutes
)

agent = Agent(
    "anthropic:claude-sonnet-4-6",
    toolsets=[dtc_server],
    system_prompt="You are a mobile app developer. Use DTC tools to build apps.",
)

async with agent:
    result = await agent.run("Build a pet adoption app with SwiftUI")

Connecting from Claude Desktop

Add to your claude_desktop_config.json:

{
  "mcpServers": {
    "dtc": {
      "command": "node",
      "args": ["/absolute/path/to/packages/appifex-dtc/packages/mcp-server/dist/index.js"],
      "env": { "HOME": "/Users/yourname" }
    }
  }
}

Connecting from Claude Code

Add to your project's .mcp.json:

{
  "mcpServers": {
    "dtc": {
      "command": "node",
      "args": ["packages/appifex-dtc/packages/mcp-server/dist/index.js"]
    }
  }
}

Connecting from the Appifex AI Technologies, Inc. Backend

Option A: Pydantic AI Agent Toolset

from pydantic_ai.mcp import MCPServerStdio

dtc_mcp = MCPServerStdio(
    "node",
    args=[str(DTC_MCP_SERVER_PATH / "dist" / "index.js")],
    env={"HOME": os.environ.get("HOME", "")},
    timeout=30,
    read_timeout=600,
)

agent = Agent("model", toolsets=[dtc_mcp, other_toolsets...])

Option B: Claude Agent SDK MCP Server

# In backend/app/claude_code/runner.py
options = ClaudeAgentOptions(
    mcp_servers={
        "appifex": appifex_mcp_server,
        "dtc": dtc_mcp_server_config,
    },
    ...
)
# Tools appear as mcp__dtc__build, mcp__dtc__validate, etc.

Available Tools (21)

Pipeline (Full Orchestration)

Tool	Description
dtc_run_pipeline	Run the entire design-to-code pipeline (design, spec, test-gen, codegen, build, validate, fix, deliver, report)

Prompt Refinement

Tool	Description
dtc_refine_prompt	Refine a vague app prompt into a detailed one. Use "ask" mode to get clarifying questions, then "enrich" mode with user answers to produce a pipeline-ready prompt
dtc_refine_feature_prompt	Refine a vague add-feature prompt into a detailed prompt with assumptions. First call returns assumptions for confirmation, second call with confirmed=true returns the enriched prompt

Add Feature

Tool	Description
dtc_add_feature	Add a feature to an existing project. Validates that a prior completed run exists at outputDir before proceeding

Design

Tool	Description
dtc_design_create	Generate a .pen design file from a text prompt
dtc_design_iterate	Apply changes to an existing .pen design file

Spec

Tool	Description
dtc_spec_extract	Extract a DesignSpec from a .pen file (deterministic, no LLM)
dtc_spec_translate	Translate a DesignSpec to a platform-specific PlatformSpec (SwiftUI or Kotlin Compose)

Test Generation

Tool	Description
dtc_test_gen_ui	Generate Maestro UI test flows from a PlatformSpec
dtc_test_gen_unit	Generate unit tests (XCTest or JUnit) from a PlatformSpec

Build

Tool	Description
dtc_build	Build a project (xcodebuild for SwiftUI, Gradle for Kotlin Compose)

Validation

Tool	Description
dtc_validate	Run all tests: Maestro UI + unit tests + optional Semgrep security scan
dtc_security	Run Semgrep OWASP security scan

Fix

Tool	Description
dtc_fix	Run the TDD fix loop (fix, build, validate, repeat until green or circuit breaker)

Deliver

Tool	Description
dtc_deliver	Git commit + push + PR creation. Auto-creates GitHub repo if needed.

Report

Tool	Description
dtc_report	Generate a pipeline report (markdown or JSON)

Provision & Submit

Tool	Description
dtc_provision_submit	Build and submit an app to TestFlight (iOS) or Play Console Internal Testing (Android). Auto-detects platform from the project directory

Analysis

Tool	Description
dtc_analyze	Scan an existing project to produce a structural inventory and navigation graph

Config & Context

Tool	Description
dtc_load_config	Load DTC configuration from ~/.dtc/config.json
dtc_load_context	Load the previous run context from a project directory
dtc_save_context	Save a run context for future resume/add-feature/refactor

Tool Details

dtc_run_pipeline

The highest-level tool. Runs the entire pipeline in non-interactive mode.

Input:

{
  "prompt": "Pet adoption app with browse, favorites, and adoption form",
  "platform": "swiftui",
  "outputDir": "/path/to/output",
  "designFile": "/optional/path/to/design.pen",
  "mode": "fresh",
  "agentType": "auto",
  "verbose": false,
  "benchmark": false,
  "resumeSessionId": "optional-session-id",
  "configDir": "/optional/path/to/.dtc",
  "baasProvider": "firebase"
}

Output:

{
  "status": "completed",
  "summary": { "allGreen": true, "totalTests": 12, "totalPassed": 12 },
  "markdown": "# Pet App\n**Status:** ALL GREEN...",
  "deliver": { "commitHash": "abc1234", "branch": "dtc/1712100000" },
  "events": ["[design] started: Creating design...", "..."]
}

dtc_build

Input:

{
  "platform": "swiftui",
  "projectDir": "/path/to/project",
  "scheme": "MyApp"
}

Output:

{
  "success": true,
  "duration": 12345,
  "errors": []
}

dtc_validate

Input:

{
  "platform": "swiftui",
  "projectDir": "/path/to/project",
  "runSecurity": true
}

Output:

{
  "allPassed": false,
  "ui": { "total": 4, "passed": 3, "failed": 1 },
  "unit": { "total": 8, "passed": 8, "failed": 0 },
  "security": { "total": 0, "passed": 0, "failed": 0, "findings": [] }
}

dtc_fix

Input:

{
  "platform": "swiftui",
  "projectDir": "/path/to/project",
  "maxAttempts": 5,
  "tokenBudget": 200000
}

Output:

{
  "status": "all_green",
  "attempts": [{ "attempt": 1, "testsBefore": {...}, "testsAfter": {...} }],
  "totalTokensUsed": 15000,
  "totalDuration": 45000
}

dtc_refine_prompt

Two-step flow: first call with mode: "ask" returns clarifying questions, then call with mode: "enrich" and user answers to get a pipeline-ready prompt.

Input (ask mode):

{
  "prompt": "todo app",
  "mode": "ask",
  "platform": "swiftui"
}

Output (ask mode):

{
  "mode": "ask",
  "prompt": "todo app",
  "completenessScore": 17,
  "questions": [
    { "id": "screens", "question": "What screens should the app have?", "category": "Screens", "required": true },
    { "id": "navigation", "question": "What navigation pattern?", "category": "Navigation", "options": ["Tab bar", "Stack", "Drawer"], "required": true }
  ],
  "hint": "Or just say \"just build it\" to skip all questions and build with sensible defaults."
}

Input (enrich mode):

{
  "prompt": "todo app",
  "mode": "enrich",
  "answers": "{\"screens\": \"home, detail, settings\", \"navigation\": \"Tab bar\"}",
  "platform": "swiftui"
}

dtc_add_feature

Input:

{
  "prompt": "Add a favorites screen with heart button on each item",
  "outputDir": "/path/to/existing/project",
  "platform": "swiftui",
  "confirmed": true
}

dtc_provision_submit

Input:

{
  "projectDir": "/path/to/project",
  "platform": "ios",
  "scheme": "App",
  "exportMethod": "app-store"
}

dtc_analyze

Input:

{
  "outputDir": "/path/to/existing/project",
  "platform": "swiftui"
}

dtc_spec_extract + dtc_spec_translate

Typically used in sequence:

1. dtc_spec_extract({ filePath: "design.pen" })
   → returns DesignSpec JSON

2. dtc_spec_translate({ specJson: <result>, platform: "swiftui" })
   → returns PlatformSpec JSON with SwiftUI types, testIds, SF Symbols

Architecture

AI Agent (Pydantic AI / Claude Desktop / Claude Code)
    |
    | MCPServerStdio (spawns as subprocess)
    v
@appifex/mcp-server (TypeScript, stdio transport)
    |
    | imports @appifex/core, @appifex/runner, @appifex/design, @appifex/spec, etc.
    v
DTC Package APIs
    |
    | Runner.exec(), Runner.readFile(), etc.
    v
Local Machine / E2B Sandbox / Remote Mac Runner

Tool handlers are separated from MCP wiring in src/tools/ — each is a plain async function that can be tested independently without the MCP SDK.

Configuration

The MCP server reads DTC config from ~/.dtc/config.json (or a custom path via the configDir parameter on most tools). Run dtc setup to configure:

• LLM provider (Anthropic, OpenAI, Google, Copilot, Claude CLI)
• Design tool (Pencil, Google Stitch, or Figma Make; plus zero-config .zip import for Stitch / Figma Make / Claude Design exports via designFile)
• Runner type (local, E2B, remote)
• Apple credentials (optional)
• Deliver config (optional)

Development

# Install deps
pnpm install

# Run tests
npx vitest run packages/mcp-server/__tests__/

# Build
pnpm --filter @appifex/mcp-server run build

Important Notes

• The server uses stdio transport — never use console.log() in tool handlers (it corrupts the JSON-RPC stream). Use console.error() for debugging.
• Long-running tools (dtc_fix, dtc_run_pipeline) can take minutes. Set read_timeout=600 or higher in MCPServerStdio.
• The dtc_run_pipeline tool runs in non-interactive mode — all TTY prompts (design review, budget continuation) are skipped.
• All tools return { content: [{ type: "text", text: "..." }], isError: boolean } following the MCP protocol.