MCPEmulate
# MCPEmulate
> This project was vibecoded.
An [MCP](https://modelcontextprotocol.io/) server that exposes CPU emulation, disassembly, and assembly as tools for LLM agents. Built on [Unicorn](https://www.unicorn-engine.org/) (emulation), [Capstone](https://www.capstone-engine.org/) (disassembly), [Keystone](https://www.keystone-engine.org/) (assembly), and [LIEF](https://lief-project.github.io/) (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+.
```bash
# 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:
```json
{
"mcpServers": {
"mcp-emulate": {
"command": "uvx",
"args": ["mcp-emulate"]
}
}
}
```
### CLI
```bash
# 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:
1. `create_emulator(arch="x86_64")` -- start a session
2. `assemble(arch="x86_64", code="mov rax, 60; syscall")` -- assemble exit syscall
3. `load_binary(session_id=..., data=..., address=0x1000, entry_point=0x1000)` -- load code
4. `hook_syscall(session_id=..., mode="stop")` -- intercept syscalls
5. `enable_trace(session_id=...)` -- start recording
6. `emulate(session_id=..., address=0x1000, count=100)` -- run
7. `get_trace(session_id=...)` -- inspect what executed
8. `get_syscall_log(session_id=...)` -- see what syscalls were attempted
9. `export_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
- **`EmulationSession` uses `__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_mode` are `Optional`** on `ArchConfig` so architectures without Keystone (RISC-V) can exist without a dummy value. The `assemble` tool checks this and returns a clear error.
- **`load_executable` writes via `uc.mem_write()` directly**, bypassing the permission check on `write_memory`. This is intentional -- binary loaders need to populate read-only segments.
- **Session serialization is versioned** (`"version": 1`) for forward compatibility.
## Development
```bash
git clone https://github.com/LabGuy94/MCPEmulate.git
cd MCPEmulate
uv venv .venv
uv pip install -e ".[dev]"
```
## Tests
```bash
# 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](https://pypi.org/project/mcp/) >= 1.18.0 | MCP protocol / FastMCP server framework |
| [unicorn](https://pypi.org/project/unicorn/) >= 2.0.0 | CPU emulation engine |
| [capstone](https://pypi.org/project/capstone/) >= 5.0.0 | Disassembly engine |
| [keystone-engine](https://pypi.org/project/keystone-engine/) >= 0.9.2 | Assembly engine |
| [lief](https://pypi.org/project/lief/) >= 0.14.0 | ELF/PE/Mach-O binary parsing |
## License
GPL-2.0-only
TDQS
Scored across 41 tools
Most tools have distinct purposes, but some overlap exists: 'hexdump' and 'read_memory' both retrieve memory data, and 'list_breakpoints'/'list_watchpoints'/'list_symbols' are similar listing operations. Descriptions help clarify differences, but an agent might occasionally misselect between closely related tools.
Tool names follow a highly consistent verb_noun pattern throughout (e.g., 'add_breakpoint', 'create_emulator', 'get_registers'). All names use snake_case with clear, descriptive verbs, making the set predictable and easy to navigate.
With 41 tools, the count is excessive for a CPU emulation server. While the domain is complex, many tools could be consolidated (e.g., multiple listing tools) or omitted without losing core functionality, leading to potential confusion and inefficiency for agents.
The toolset provides comprehensive coverage for CPU emulation, including session management, memory operations, debugging (breakpoints/watchpoints), tracing, syscall handling, and state import/export. All essential CRUD and lifecycle operations are present, with no obvious gaps in the domain.