VICE C64 Emulator MCP Server
Enables autonomous debugging of Commodore 64 programs through the VICE emulator with semantic interpretation of C64-specific data structures, memory layouts, VIC-II states, and PETSCII encoding for AI-assisted 6502 assembly debugging.
README
vice-mcp
A Model Context Protocol (MCP) server for autonomous C64 debugging via the VICE emulator.
What is this?
vice-mcp bridges AI agents to the VICE Commodore 64 emulator, enabling autonomous debugging of 6502 assembly programs. Unlike raw protocol wrappers, it provides a semantic layer that interprets C64-specific data structures and returns meaningful, actionable information.
Why this exists:
- AI agents need more than hex dumps—they need interpreted data with context
- Debugging C64 code requires understanding VIC-II banks, PETSCII encoding, sprite pointers, and memory layouts
- Every response includes hints suggesting next steps and related tools
Key differentiators:
- Semantic output:
readScreenreturns text, not screen codes.readVicStateexplains graphics modes, not register bits. - Actionable hints: Every response suggests what to do next
- Cross-references: Tools point to related tools for common workflows
- Agent-friendly errors: Clear error codes and recovery suggestions
Prerequisites
- Node.js 18 or later
- VICE emulator with binary monitor enabled
Starting VICE with Binary Monitor
# x64sc is the accurate C64 emulator (recommended)
x64sc -binarymonitor -binarymonitoraddress ip4://127.0.0.1:6502
# Or with x64 (faster, less accurate)
x64 -binarymonitor -binarymonitoraddress ip4://127.0.0.1:6502
The binary monitor listens on port 6502 by default.
Installation
From npm (when published)
npx @simen/vice-mcp
From GitHub
npx github:simen/vice-mcp
Local Development
git clone https://github.com/simen/vice-mcp.git
cd vice-mcp
npm install
npm run build
npm start
Claude Code Installation
The quickest way to get started with Claude Code:
1. Start VICE with binary monitor:
x64sc -binarymonitor -binarymonitoraddress ip4://127.0.0.1:6502
2. Add the MCP server:
claude mcp add vice-mcp -- npx github:simen/vice-mcp
3. Restart Claude Code to load the new MCP server.
That's it! You can now ask Claude Code to debug your C64 programs.
Manual Configuration
Alternatively, add to ~/.claude/claude_desktop_config.json:
{
"mcpServers": {
"vice-mcp": {
"command": "npx",
"args": ["github:simen/vice-mcp"]
}
}
}
Configuration
Add to your MCP client configuration (e.g., Claude Desktop, Cursor, or custom agent):
{
"mcpServers": {
"vice": {
"command": "npx",
"args": ["@simen/vice-mcp"]
}
}
}
Or for local development:
{
"mcpServers": {
"vice": {
"command": "node",
"args": ["/path/to/vice-mcp/dist/index.js"]
}
}
}
Tool Reference
Connection & Status
| Tool | Description |
|---|---|
connect |
Connect to VICE (default: 127.0.0.1:6502) |
disconnect |
Disconnect from VICE |
status |
Get connection state and emulation status |
Memory Operations
| Tool | Description |
|---|---|
readMemory |
Read raw bytes with hex dump and ASCII |
writeMemory |
Write bytes to memory |
CPU & Execution
| Tool | Description |
|---|---|
getRegisters |
Get A, X, Y, SP, PC, and flags (interpreted) |
step |
Single-step execution (with step-over option) |
continue |
Resume execution |
reset |
Soft or hard reset |
runTo |
Run until specific address (temporary breakpoint) |
disassemble |
Disassemble 6502 code with KERNAL labels |
Breakpoints & Watchpoints
| Tool | Description |
|---|---|
setBreakpoint |
Set execution breakpoint |
deleteBreakpoint |
Remove breakpoint or watchpoint |
listBreakpoints |
List all breakpoints |
toggleBreakpoint |
Enable/disable breakpoint |
setWatchpoint |
Set memory read/write watchpoint |
listWatchpoints |
List all watchpoints |
Semantic Layer (Interpreted C64 Data)
| Tool | Description |
|---|---|
readScreen |
Get screen as text (PETSCII decoded) with summary mode |
readColorRam |
Get color RAM with color names and usage stats |
readVicState |
Full VIC-II state: graphics mode, colors, banks, sprites |
readSprites |
All 8 sprites: position, visibility, colors, pointers |
Visual Feedback
| Tool | Description |
|---|---|
screenshot |
Capture display buffer with palette |
renderScreen |
ASCII art rendering of display |
State Management
| Tool | Description |
|---|---|
saveSnapshot |
Save complete machine state to file |
loadSnapshot |
Load machine state from file |
loadProgram |
Load and optionally run PRG/D64/T64 files |
Example Usage
Basic Debugging Session
1. connect() → Establish connection
2. loadProgram("game.prg") → Load the program
3. setBreakpoint(0x0810) → Break at main loop
4. continue() → Run until breakpoint
5. getRegisters() → Check CPU state
6. readScreen() → See what's on screen
7. step(count: 5) → Execute 5 instructions
8. disassemble() → See code at current PC
Debugging Sprite Issues
1. readVicState() → Check sprite enable bits
2. readSprites(enabledOnly: true) → Get enabled sprite details
→ Response includes visibility check and position analysis
3. If sprite not visible, hint tells you why (off-screen, wrong bank, etc.)
Memory Watchpoint Workflow
1. setWatchpoint(startAddress: 0x0400, type: "store")
→ Watch for writes to screen RAM
2. continue()
→ Execution stops when something writes to screen
3. getRegisters()
→ See PC to find the code that wrote
4. disassemble()
→ Understand what the code is doing
State Checkpoint Pattern
1. saveSnapshot("before-test.vsf") → Save state
2. [Make changes, test things]
3. loadSnapshot("before-test.vsf") → Restore to known state
Response Format
All responses include:
- Structured data with
valueandhexrepresentations _metablock with connection statehintfield with contextual next steps
Example getRegisters response:
{
"a": { "value": 65, "hex": "$41" },
"x": { "value": 0, "hex": "$00" },
"y": { "value": 0, "hex": "$00" },
"sp": { "value": 243, "hex": "$f3", "stackTop": "$01f3" },
"pc": { "value": 2049, "hex": "$0801" },
"flags": {
"negative": false,
"overflow": false,
"zero": false,
"carry": false,
"string": "nv-bdizc"
},
"hint": "CPU state looks normal",
"_meta": {
"connected": true,
"running": false,
"host": "127.0.0.1",
"port": 6502
}
}
Architecture Overview
┌─────────────────────────────────────────────────────────┐
│ MCP Client (Agent) │
└─────────────────────────────────────────────────────────┘
│
│ MCP Protocol (stdio)
▼
┌─────────────────────────────────────────────────────────┐
│ src/index.ts │
│ (MCP Server) │
│ ┌─────────────────────────────────────────────────┐ │
│ │ Tool Handlers (24 tools) │ │
│ │ • Connection: connect, disconnect, status │ │
│ │ • Memory: readMemory, writeMemory │ │
│ │ • CPU: getRegisters, step, continue, reset │ │
│ │ • Breakpoints: set, delete, list, toggle │ │
│ │ • Watchpoints: set, list │ │
│ │ • Semantic: readScreen, readVicState, etc. │ │
│ │ • Visual: screenshot, renderScreen │ │
│ │ • State: saveSnapshot, loadSnapshot, loadPrg │ │
│ └─────────────────────────────────────────────────┘ │
└─────────────────────────────────────────────────────────┘
│
│ Uses
▼
┌─────────────────────────────────────────────────────────┐
│ src/protocol/client.ts │
│ (ViceClient) │
│ • TCP socket connection to VICE │
│ • Binary protocol encoding/decoding │
│ • Request/response correlation │
│ • Checkpoint (breakpoint/watchpoint) tracking │
└─────────────────────────────────────────────────────────┘
│
│ TCP Socket
▼
┌─────────────────────────────────────────────────────────┐
│ VICE Binary Monitor │
│ (Port 6502) │
└─────────────────────────────────────────────────────────┘
Key Files
| File | Purpose |
|---|---|
src/index.ts |
MCP server, tool definitions, semantic layer |
src/protocol/client.ts |
VICE binary monitor client |
src/protocol/types.ts |
Protocol constants and types |
src/utils/c64.ts |
C64 utilities (PETSCII, colors, VIC banks) |
src/utils/disasm.ts |
6502 disassembler with all addressing modes |
Design Principles
- Semantic over raw: Return interpreted data, not just bytes
- Hints everywhere: Every response suggests next actions
- Cross-references: Tools reference related tools
- Fail informatively: Errors explain what went wrong and how to fix it
- Agent-first: Designed for autonomous operation, not human CLI use
Protocol Reference
vice-mcp implements the VICE Binary Monitor Protocol. Key commands used:
| Code | Command | Purpose |
|---|---|---|
| 0x01 | MemoryGet | Read memory |
| 0x02 | MemorySet | Write memory |
| 0x12 | CheckpointSet | Create breakpoint/watchpoint |
| 0x13 | CheckpointDelete | Remove checkpoint |
| 0x15 | CheckpointToggle | Enable/disable checkpoint |
| 0x31 | RegistersGet | Read CPU registers |
| 0x41 | Dump | Save snapshot |
| 0x42 | Undump | Load snapshot |
| 0x81 | Continue | Resume execution |
| 0x82 | Step | Single-step |
| 0x84 | DisplayGet | Capture screen |
| 0x91 | PaletteGet | Get color palette |
| 0xdd | AutoStart | Load and run program |
License
MIT
推荐服务器
Baidu Map
百度地图核心API现已全面兼容MCP协议,是国内首家兼容MCP协议的地图服务商。
Playwright MCP Server
一个模型上下文协议服务器,它使大型语言模型能够通过结构化的可访问性快照与网页进行交互,而无需视觉模型或屏幕截图。
Magic Component Platform (MCP)
一个由人工智能驱动的工具,可以从自然语言描述生成现代化的用户界面组件,并与流行的集成开发环境(IDE)集成,从而简化用户界面开发流程。
Audiense Insights MCP Server
通过模型上下文协议启用与 Audiense Insights 账户的交互,从而促进营销洞察和受众数据的提取和分析,包括人口统计信息、行为和影响者互动。
VeyraX
一个单一的 MCP 工具,连接你所有喜爱的工具:Gmail、日历以及其他 40 多个工具。
graphlit-mcp-server
模型上下文协议 (MCP) 服务器实现了 MCP 客户端与 Graphlit 服务之间的集成。 除了网络爬取之外,还可以将任何内容(从 Slack 到 Gmail 再到播客订阅源)导入到 Graphlit 项目中,然后从 MCP 客户端检索相关内容。
Kagi MCP Server
一个 MCP 服务器,集成了 Kagi 搜索功能和 Claude AI,使 Claude 能够在回答需要最新信息的问题时执行实时网络搜索。
e2b-mcp-server
使用 MCP 通过 e2b 运行代码。
Neon MCP Server
用于与 Neon 管理 API 和数据库交互的 MCP 服务器
Exa MCP Server
模型上下文协议(MCP)服务器允许像 Claude 这样的 AI 助手使用 Exa AI 搜索 API 进行网络搜索。这种设置允许 AI 模型以安全和受控的方式获取实时的网络信息。