sbl-debugger
Enables AI assistants to debug ARM Cortex-M targets via GDB and OpenOCD, supporting attach, breakpoints, stepping, register/memory inspection, and SVD peripheral decoding.
README
<!-- mcp-name: io.github.soundbytelabs/debugger -->
sbl-debugger
Embedded debug MCP server for ARM Cortex-M targets. Gives AI coding assistants direct control of GDB and OpenOCD — attach to hardware, set breakpoints, step through code, inspect registers and memory, all within a conversation.
Part of the Sound Byte Labs embedded tooling suite, alongside sbl-probe for serial I/O.
Installation
Create a virtual environment and install the package:
python3 -m venv .venv
source .venv/bin/activate
# Install
pip install -e .
# Or with test dependencies
pip install -e ".[dev]"
System Requirements
gdb-multiarch(GDB with ARM target support)openocd(0.12.0+ recommended)- SWD debug probe (ST-LINK, CMSIS-DAP, etc.)
On Debian/Ubuntu/Raspberry Pi OS:
sudo apt install gdb-multiarch openocd
MCP Configuration
Register the server in your MCP client's config. For most clients, add to .mcp.json in your project root:
{
"mcpServers": {
"sbl-debugger": {
"type": "stdio",
"command": "/absolute/path/to/.venv/bin/python",
"args": ["-m", "sbl_debugger"]
}
}
}
Important: Use the absolute path to the Python binary inside your virtual environment. For example:
/home/you/sbl-debugger-mcp/.venv/bin/python
Restart your MCP client and the tools are available immediately.
Tools
Session Management
| Tool | Description |
|---|---|
debug_attach |
Attach to a target — launches OpenOCD + GDB, connects via SWD |
debug_detach |
Cleanly shut down a debug session |
debug_sessions |
List all active debug sessions |
debug_status |
Get current target state (halted/running, stop reason, current frame) |
debug_targets |
List available predefined target profiles |
Execution Control
| Tool | Description |
|---|---|
halt |
Halt a running target |
continue_execution |
Resume execution |
wait_for_halt |
Block until target stops (breakpoint hit, etc.) |
step |
Step source lines (into functions) |
step_over |
Step source lines (over function calls) |
step_out |
Step out of current function |
step_instruction |
Step machine instructions |
run_to |
Run to a location (sets temporary breakpoint + continues) |
reset |
Reset the target (halt or run after reset) |
Inspection
| Tool | Description |
|---|---|
read_registers |
Read CPU registers (all core regs or specific subset) |
write_register |
Write a CPU register |
read_memory |
Read memory (hex, u8, u16, u32 formats) |
write_memory |
Write to memory |
backtrace |
Get the call stack |
read_locals |
List local variables in current frame |
print_expr |
Evaluate a C/C++ expression in target context |
disassemble |
Disassemble at an address or current PC |
Breakpoints
| Tool | Description |
|---|---|
breakpoint_set |
Set breakpoint by function name, file:line, or address |
breakpoint_delete |
Delete a breakpoint |
breakpoint_list |
List all breakpoints and watchpoints |
watchpoint_set |
Set a hardware data watchpoint (write/read/access) |
Peripheral Registers (SVD)
| Tool | Description |
|---|---|
list_peripherals |
List SVD peripherals with base addresses and register counts |
list_registers |
Show all registers and bitfield definitions for a peripheral |
read_peripheral_register |
Read a register from hardware and decode all bitfields |
read_peripheral |
Read all registers of a peripheral with decoded bitfields |
Requires
cecrops(optional dependency) andSBL_HW_PATHenvironment variable pointing to sbl-hardware. Install with:pip install -e ".[svd]"
Snapshot & Advanced
| Tool | Description |
|---|---|
debug_snapshot |
Full target state in one call (frame, registers, backtrace, locals, source) |
load |
Flash firmware to the target |
monitor |
Send raw OpenOCD monitor commands |
Target Profiles
Predefined profiles eliminate the need to remember OpenOCD configs:
| Profile | Hardware | Debug Probe |
|---|---|---|
daisy |
Electrosmith Daisy Seed (STM32H750, Cortex-M7) | ST-LINK |
pico |
Raspberry Pi Pico (RP2040, Cortex-M0+) | CMSIS-DAP Debug Probe |
pico2 |
Raspberry Pi Pico 2 (RP2350, Cortex-M33) | CMSIS-DAP Debug Probe |
custom |
Any target | Provide interface and target_cfg explicitly |
Custom targets work with any OpenOCD-supported hardware:
debug_attach(target="custom", interface="jlink.cfg", target_cfg="stm32f4x.cfg", elf="firmware.elf")
Architecture
sbl_debugger/
├── server.py # FastMCP server, tool wiring
├── targets.py # Target profiles (daisy, pico, pico2)
├── session/
│ ├── manager.py # Thread-safe session registry
│ └── session.py # DebugSession (owns OpenOCD + GDB)
├── process/
│ ├── openocd.py # OpenOCD subprocess lifecycle
│ └── ports.py # GDB server port allocation
├── bridge/
│ ├── mi.py # GDB/MI wrapper (pygdbmi + lock)
│ └── types.py # FrameInfo, StopEvent, MiResult
├── svd/
│ ├── peripheral_db.py # SVD lookup/decode (wraps cecrops Device)
│ └── loader.py # SBL_HW_PATH resolution, cecrops import guard
└── tools/
├── session.py # attach, detach, sessions, status, targets
├── execution.py # halt, continue, step, reset
├── inspection.py # registers, memory, backtrace, locals
├── breakpoints.py # breakpoint/watchpoint management
├── snapshot.py # combined state dump
├── advanced.py # load (flash), monitor (raw OpenOCD)
└── peripheral.py # SVD peripheral register decoding
Key design decisions:
- Managed subprocesses — OpenOCD and GDB are launched, monitored, and cleaned up by the server. One
debug_attachcall does everything. - GDB/MI via pygdbmi — structured command/response interface, no string parsing of GDB CLI output
- Single command lock — one MI command at a time prevents response interleaving
- Explicit polling —
wait_for_haltanddebug_statuscheck target state on demand, matching MCP's request/response model - Error dicts, not exceptions — tools return
{"error": "..."}instead of crashing the server
Running Tests
pytest # 275 tests (all mocked, no hardware needed)
pytest -v # verbose
pytest tests/test_tools.py # just tool tests
Dependencies
mcp— Official Python MCP SDK (FastMCP)pygdbmi— GDB Machine Interface protocolcecrops— SVD register definitions (optional, for peripheral tools)- Python >= 3.11
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。