MCPEmulate
An MCP server that provides CPU emulation, disassembly, and assembly tools for LLM agents across multiple architectures including x86, ARM, and RISC-V. It enables agents to manage isolated emulation sessions, perform memory analysis, hook syscalls, and trace execution through a standard tool interface.
README
MCPEmulate
This project was vibecoded.
An MCP server that exposes CPU emulation, disassembly, and assembly as tools for LLM agents. Built on Unicorn (emulation), Capstone (disassembly), Keystone (assembly), and LIEF (binary parsing).
Agents can create isolated emulation sessions, load code or full executables, set breakpoints, hook syscalls, step through instructions, inspect memory and registers, and diff execution traces -- all through the standard MCP tool interface.
Supported Architectures
| Architecture | Emulation | Disassembly | Assembly | Syscall Hooking |
|---|---|---|---|---|
| x86 (32-bit) | Yes | Yes | Yes | int 0x80 |
| x86-64 | Yes | Yes | Yes | syscall |
| ARM (32-bit) | Yes | Yes | Yes | svc 0 |
| AArch64 | Yes | Yes | Yes | svc 0 |
| MIPS32 (LE) | Yes | Yes | Yes | syscall |
| MIPS32 (BE) | Yes | Yes | Yes | syscall |
| RISC-V 32 | Yes | Yes | No | ecall |
| RISC-V 64 | Yes | Yes | No | ecall |
RISC-V architectures lack a Keystone backend, so the assemble tool returns an error for them. Disassembly and emulation work normally.
Install
Requires Python 3.10+.
# Run directly (no install needed)
uvx mcp-emulate
# Or install globally
uv pip install mcp-emulate
Usage
Claude Desktop / MCP Client
Add to your MCP client configuration:
{
"mcpServers": {
"mcp-emulate": {
"command": "uvx",
"args": ["mcp-emulate"]
}
}
}
CLI
# Default: stdio transport (for MCP clients)
mcp-emulate
# SSE transport (network, for web-based clients)
mcp-emulate --transport sse
# Streamable HTTP transport (newer MCP protocol)
mcp-emulate --transport streamable-http
Tools (41)
Session Management
| Tool | Description |
|---|---|
create_emulator |
Create a new emulation session for a given architecture |
destroy_emulator |
Destroy a session and free resources |
export_session |
Export full session state (memory, registers, breakpoints, symbols) to JSON |
import_session |
Create a new session and restore state from a previous export |
Memory
| Tool | Description |
|---|---|
map_memory |
Map a memory region with specified permissions (r/w/x) |
write_memory |
Write hex or base64 data to memory |
read_memory |
Read memory as hex or base64 |
list_regions |
List all mapped regions |
hexdump |
Formatted hex dump (up to 4KB) with ASCII sidebar |
search_memory |
Search for byte patterns across mapped memory |
snapshot_memory |
Capture all memory content under a named label |
diff_memory |
Compare two snapshots and return changed byte ranges |
memory_map |
/proc/self/maps-style layout with gaps and symbol annotations |
Registers
| Tool | Description |
|---|---|
set_registers |
Write one or more registers |
get_registers |
Read registers (specific or all) |
get_stack |
Read stack entries from SP, resolving values against symbols |
Execution
| Tool | Description |
|---|---|
emulate |
Run emulation with stop address, instruction count, or timeout |
step |
Execute a single instruction with full disassembly |
add_breakpoint |
Set a breakpoint, optionally with a register condition |
remove_breakpoint |
Remove a breakpoint |
list_breakpoints |
List all breakpoints with their conditions |
save_context |
Save a register snapshot under a label |
restore_context |
Restore registers from a saved snapshot |
Breakpoint Conditions
Conditional breakpoints accept expressions like:
eax == 42
rax > 0x1000 and rcx != 0
r0 == 0 or r1 & 0xff
Supported operators: ==, !=, >, <, >=, <=, &. Connectives: and, or.
Syscall Hooking
| Tool | Description |
|---|---|
hook_syscall |
Install a syscall hook (skip to log and continue, stop to halt) |
unhook_syscall |
Remove the syscall hook |
get_syscall_log |
Retrieve logged syscall invocations with pagination |
Each logged entry includes the syscall number, argument register values, and PC. The hook is architecture-aware -- it intercepts int 0x80 on x86_32, syscall on x86_64, svc 0 on ARM/AArch64, syscall on MIPS, and ecall on RISC-V.
Watchpoints
| Tool | Description |
|---|---|
add_watchpoint |
Watch a memory address for read, write, or both |
remove_watchpoint |
Remove a watchpoint |
list_watchpoints |
List all active watchpoints |
Tracing
| Tool | Description |
|---|---|
enable_trace |
Start recording executed instructions |
disable_trace |
Stop recording (log is preserved) |
get_trace |
Retrieve trace entries with disassembly and pagination |
save_trace |
Save the current trace log under a named label |
diff_trace |
Compare two saved traces instruction-by-instruction |
Trace diff reports the common prefix length, the divergence point, and up to 50 differing entries with full disassembly.
Symbols
| Tool | Description |
|---|---|
add_symbol |
Associate a name with an address |
remove_symbol |
Remove a symbol |
list_symbols |
List all symbols |
Symbols are used to annotate stack entries, trace output, memory maps, and step results.
Loading
| Tool | Description |
|---|---|
load_binary |
Load raw machine code at an address, auto-mapping memory |
load_executable |
Load an ELF, PE, or Mach-O binary with correct segment permissions, entry point, and symbols |
assemble |
Assemble instructions to machine code (standalone, no session) |
disassemble |
Disassemble machine code to instructions (standalone, no session) |
load_executable uses LIEF for format detection. It maps each loadable segment with the correct permissions, sets PC to the entry point, and registers exported symbols automatically.
Example Workflow
A typical agent interaction:
create_emulator(arch="x86_64")-- start a sessionassemble(arch="x86_64", code="mov rax, 60; syscall")-- assemble exit syscallload_binary(session_id=..., data=..., address=0x1000, entry_point=0x1000)-- load codehook_syscall(session_id=..., mode="stop")-- intercept syscallsenable_trace(session_id=...)-- start recordingemulate(session_id=..., address=0x1000, count=100)-- runget_trace(session_id=...)-- inspect what executedget_syscall_log(session_id=...)-- see what syscalls were attemptedexport_session(session_id=...)-- save state for later
Architecture
src/mcp_emulate/
architectures.py Architecture configs, register maps, syscall conventions
session.py EmulationSession (Unicorn wrapper), SessionManager
server.py 41 MCP tool handlers via FastMCP
tests/
test_emulate.py 132 pytest unit tests
test_server_integration.py 112 checks over JSON-RPC (23 phases)
Key Design Decisions
EmulationSessionuses__slots__for memory efficiency and to catch typos. Every new attribute must be declared.- Breakpoints are
dict[int, str | None](address to optional condition), not a set. This supports conditional breakpoints while keeping the same lookup semantics. - Syscall conventions are data, not code. A frozen dataclass per architecture describes the hook type, interrupt number filter, register names for nr/args/return. The hooking logic is generic.
ks_arch/ks_modeareOptionalonArchConfigso architectures without Keystone (RISC-V) can exist without a dummy value. Theassembletool checks this and returns a clear error.load_executablewrites viauc.mem_write()directly, bypassing the permission check onwrite_memory. This is intentional -- binary loaders need to populate read-only segments.- Session serialization is versioned (
"version": 1) for forward compatibility.
Development
git clone https://github.com/LabGuy94/MCPEmulate.git
cd MCPEmulate
uv venv .venv
uv pip install -e ".[dev]"
Tests
# Unit tests (132 tests, ~1s)
uv run pytest tests/test_emulate.py -v
# Integration tests (112 checks over JSON-RPC subprocess, ~30s)
uv run python tests/test_server_integration.py
# Both
uv run pytest tests/ -v && uv run python tests/test_server_integration.py
Dependencies
| Package | Purpose |
|---|---|
| mcp >= 1.18.0 | MCP protocol / FastMCP server framework |
| unicorn >= 2.0.0 | CPU emulation engine |
| capstone >= 5.0.0 | Disassembly engine |
| keystone-engine >= 0.9.2 | Assembly engine |
| lief >= 0.14.0 | ELF/PE/Mach-O binary parsing |
License
GPL-2.0-only
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。