MCP 服务器

dbgprobe-mcp-server

Stateful MCP server for driving debug probes (J-Link) to flash, debug, and inspect embedded targets. Enables AI agents to perform flash, memory, breakpoint, and ELF/SVD-aware operations conversationally.

README

Debug Probe MCP Server

A stateful debug probe Model Context Protocol (MCP) server for developer tooling and AI agents. Works out of the box with Claude Code, VS Code with Copilot, and any MCP-compatible runtime. Communicates over stdio and drives on-chip debug probes (J-Link first, OpenOCD and pyOCD planned) to flash, debug, and inspect embedded targets.

Example: Let Claude Code list attached J-Link probes, connect to your nRF52840, flash a new firmware, read memory, and reset the target — all conversationally.

Demo

Video walkthrough — connecting to a J-Link probe, flashing firmware, loading ELF and SVD for symbol-aware debugging, RTT logging, and breakpoints.

Why this exists

If you've ever typed J-Link Commander commands by hand, copy-pasted memory addresses between a datasheet and a terminal, re-flashed the same firmware 20 times during a debug session, and juggled multiple tool windows — this is for you.

You have a microcontroller on a debug probe. You want an AI agent to interact with it — connect, flash firmware, read/write memory, reset, halt, resume. This server makes that possible.

It gives any MCP-compatible agent a full set of debug probe tools. The agent calls these tools, gets structured JSON back, and reasons about what to do next — without you manually driving JLinkExe for every operation.

What agents can do with it:

Flash and iterate — build firmware, flash it, reset, check behavior — all in one conversation
Inspect memory — read peripheral registers, check RAM contents, verify flash writes
Debug interactively — halt, step, set breakpoints, inspect state, resume
Automate test flows — flash → reset → read output → validate
Multi-probe setups — connect to multiple probes simultaneously, each with its own session

Who is this for?

Embedded engineers — faster iteration: flash, debug, inspect memory conversationally
Hobbyists and makers — interact with microcontrollers without learning JLinkExe command syntax
QA and test engineers — automated flash-and-test sequences across multiple boards
Researchers — systematic exploration of embedded systems, register inspection

Quickstart (Claude Code)

pip install dbgprobe-mcp-server

# Register the MCP server with Claude Code
claude mcp add dbgprobe -- dbgprobe_mcp

# Or with explicit J-Link path
claude mcp add dbgprobe \
  -e DBGPROBE_JLINK_PATH=/Applications/SEGGER/JLink/JLinkExe \
  -- dbgprobe_mcp

Then in Claude Code, try:

"List attached debug probes, connect to the J-Link, and read 16 bytes from address 0x20000000."

Supported backends

Backend	Status	Probe hardware
J-Link	Working (v0)	SEGGER J-Link (EDU, EDU Mini, PLUS, PRO, etc.)
OpenOCD	Planned	ST-Link, CMSIS-DAP, and many others
pyOCD	Planned	CMSIS-DAP, ST-Link, J-Link (via pyOCD)

The server is backend-agnostic — tool names (dbgprobe.*) stay the same regardless of which probe you use.

J-Link requirements

Install the SEGGER J-Link Software Pack. The server auto-detects JLinkExe on PATH or in common install locations:

macOS: /Applications/SEGGER/JLink/
Linux: /opt/SEGGER/JLink/, /usr/bin/
Windows: C:\Program Files\SEGGER\JLink\

Or set DBGPROBE_JLINK_PATH to point to the executable directly.

This project is not affiliated with or sponsored by SEGGER. J-Link is used because it's widely available and well-supported.

Tools

Category	Tools
Probe	`dbgprobe.probes.list`, `dbgprobe.connect`, `dbgprobe.erase`, `dbgprobe.disconnect`, `dbgprobe.reset`, `dbgprobe.halt`, `dbgprobe.go`, `dbgprobe.step`, `dbgprobe.status`, `dbgprobe.flash`, `dbgprobe.mem.read`, `dbgprobe.mem.write`, `dbgprobe.breakpoint.set`, `dbgprobe.breakpoint.clear`, `dbgprobe.breakpoint.list`
Introspection	`dbgprobe.connections.list`
ELF	`dbgprobe.elf.attach`, `dbgprobe.elf.info`, `dbgprobe.elf.lookup`, `dbgprobe.elf.symbols`
SVD	`dbgprobe.svd.attach`, `dbgprobe.svd.info`, `dbgprobe.svd.read`, `dbgprobe.svd.write`, `dbgprobe.svd.set_field`, `dbgprobe.svd.update_fields`, `dbgprobe.svd.list_peripherals`, `dbgprobe.svd.list_registers`, `dbgprobe.svd.list_fields`, `dbgprobe.svd.describe`
RTT	`dbgprobe.rtt.start`, `dbgprobe.rtt.stop`, `dbgprobe.rtt.read`, `dbgprobe.rtt.write`, `dbgprobe.rtt.status`
Plugins	`dbgprobe.plugin.list`, `dbgprobe.plugin.template`, `dbgprobe.plugin.load`, `dbgprobe.plugin.reload`
Tracing	`dbgprobe.trace.status`, `dbgprobe.trace.tail`

See docs/tools.md for full schemas and examples.

Install (development)

# Editable install from repo root
pip install -e ".[test]"

# Or with uv
uv pip install -e ".[test]"

MCP is a protocol — this server works with any MCP-compatible client. Below are setup instructions for the most common ones.

Add to Claude Code

# Standard setup
claude mcp add dbgprobe -- dbgprobe_mcp

# With default target device
claude mcp add dbgprobe \
  -e DBGPROBE_JLINK_DEVICE=nRF52840_xxAA \
  -- dbgprobe_mcp

# Debug logging
claude mcp add dbgprobe -e DBGPROBE_MCP_LOG_LEVEL=DEBUG -- dbgprobe_mcp

Add to VS Code / Copilot

Add to your project's .vscode/mcp.json (or create it):

{
  "servers": {
    "dbgprobe": {
      "type": "stdio",
      "command": "dbgprobe_mcp",
      "args": [],
      "env": {
        "DBGPROBE_JLINK_DEVICE": "nRF52840_xxAA"
      }
    }
  }
}

Adjust env to match your target — set DBGPROBE_JLINK_DEVICE to your chip, or remove it to specify the device at connect time.

Add to Cursor

Add to your project's .cursor/mcp.json (or create it). Cursor does not support dots in tool names, so DBGPROBE_MCP_TOOL_SEPARATOR must be set to _:

{
  "mcpServers": {
    "dbgprobe": {
      "command": "dbgprobe_mcp",
      "args": [],
      "env": {
        "DBGPROBE_JLINK_DEVICE": "nRF52840_xxAA",
        "DBGPROBE_MCP_TOOL_SEPARATOR": "_"
      }
    }
  }
}

Environment variables

Server

Variable	Default	Description
`DBGPROBE_BACKEND`	`jlink`	Debug probe backend. Future: `openocd`, `pyocd`.
`DBGPROBE_MCP_LOG_LEVEL`	`WARNING`	Python log level (`DEBUG`, `INFO`, `WARNING`, `ERROR`). Logs go to stderr.
`DBGPROBE_MCP_TRACE`	enabled	JSONL tracing of every tool call. Set to `0`, `false`, or `no` to disable.
`DBGPROBE_MCP_TRACE_PAYLOADS`	disabled	Include memory data payloads in traced args (stripped by default).
`DBGPROBE_MCP_TRACE_MAX_BYTES`	`16384`	Max payload chars before truncation (only when `TRACE_PAYLOADS` is on).
`DBGPROBE_MCP_TOOL_SEPARATOR`	`.`	Character used to separate tool name segments. Set to `_` for MCP clients that reject dots in tool names (e.g. Cursor).
`DBGPROBE_MCP_PLUGINS`	disabled	Plugin policy: `all` or comma-separated plugin names (e.g. `nrf52,stm32`).

J-Link backend

Variable	Default	Description
`DBGPROBE_JLINK_PATH`	auto-detect	Explicit path to `JLinkExe` (or `JLink.exe` on Windows).
`DBGPROBE_JLINK_GDBSERVER_PATH`	auto-detect	Explicit path to `JLinkGDBServerCLExe`.
`DBGPROBE_JLINK_DEVICE`	(none)	Default target device string (e.g. `nRF52840_xxAA`). Can be overridden per-session.
`DBGPROBE_INTERFACE`	`SWD`	Debug interface: `SWD` or `JTAG`.
`DBGPROBE_SPEED_KHZ`	`4000`	Interface clock speed in kHz.
`DBGPROBE_GDB_TRACE`	disabled	Log all GDB RSP packets to file. Set to `1`, `true`, or `yes` to enable.
`DBGPROBE_GDB_TRACE_FILE`	`/tmp/gdb_trace.log`	Path for GDB RSP trace log (only when `GDB_TRACE` is on).

ELF Support

Attach an ELF file to a session to enable symbol-aware debugging:

Symbol lookup — resolve function names to addresses and vice versa
Breakpoints by name — breakpoint.set(symbol="main") instead of raw addresses
Auto-enriched responses — status, step, and halt include symbol + symbol_offset when an ELF is attached
Flash integration — flashing an .elf auto-attaches it; flashing .hex/.bin auto-reloads a previously attached ELF; sibling .elf files are suggested via hints

> "Attach the ELF, set a breakpoint on main, run, and show me where it halted."

The agent calls elf.attach, breakpoint.set(symbol="main"), go, then status — and gets back "halted at main+0" instead of a raw hex address.

SVD Support

Attach an SVD (System View Description) file to a session to enable register-level peripheral access:

Named register reads — svd.read("GPIO.OUT") returns the raw value and all decoded fields with enum names
Field-level reads — svd.read("GPIO.PIN_CNF[3].PULL") returns the field value and enum name ("PullUp")
Safe field writes — svd.set_field("GPIO.PIN_CNF[3].PULL", "PullUp") does read-modify-write
Batch field updates — svd.update_fields("GPIO.PIN_CNF[3]", {"DIR": "Output", "PULL": "PullUp"}) — one read, one write
Raw register writes — svd.write("GPIO.OUT", 0x01) — full register, no RMW
Discovery — list peripherals, registers, fields; describe with enums
Auto-decode on mem.read — when an SVD is attached and mem.read hits a known register address, the response includes decoded fields

> "Attach the SVD, read GPIO.PIN_CNF[3], and set PULL to PullUp."

The agent calls svd.attach, svd.read("GPIO.PIN_CNF[3]"), then svd.set_field("GPIO.PIN_CNF[3].PULL", "PullUp") — and gets back decoded field values instead of raw hex.

RTT (Real-Time Transfer)

Start, stop, read, and write to SEGGER RTT channels. The agent can stream target log output and send data to the device — useful for debugging firmware that prints over RTT instead of UART.

> "Start RTT and show me the output."

For repetitive flows — read device ID, run a self-test sequence, validate calibration — you can package them as plugins: Python modules that expose custom tools the agent can call directly. Enable with DBGPROBE_MCP_PLUGINS=all. See the nrf52_info example.

Tracing

Every tool call is traced to .dbgprobe_mcp/traces/trace.jsonl and an in-memory ring buffer (last 2000 events). Tracing is on by default — set DBGPROBE_MCP_TRACE=0 to disable.

Use dbgprobe.trace.status and dbgprobe.trace.tail to inspect the trace without reading the file directly.

Try without an agent

You can test the server interactively using the MCP Inspector:

npx @modelcontextprotocol/inspector python -m dbgprobe_mcp_server

Roadmap / TODO

[ ] OpenOCD backend — support ST-Link, CMSIS-DAP, and other probes via OpenOCD subprocess
[ ] pyOCD backend — native Python probe access via pyOCD library
[ ] Multi-core support — target specific cores on multi-core SoCs
[ ] Cortex-A/R support — ARM-mode breakpoints (kind=4); currently Thumb-only (Cortex-M)

Known limitations

Single-client only. The server handles one MCP session at a time (stdio transport).
RTT channel 0 only. RTT support is limited to channel 0 (terminal). Multi-channel RTT is a future enhancement.
Flash clears breakpoints. Flashing new firmware invalidates breakpoints (the code at those addresses may have changed). The session stays alive but breakpoints are cleared.
Cortex-M only. Breakpoints use Thumb-mode (kind=2). Cortex-A/R targets (ARM-mode, kind=4) are not yet supported.
Instruction-level step only. dbgprobe.step single-steps one CPU instruction. Source-level stepping (step into/over/out) is not supported.

Safety

This server connects an AI agent to real debug hardware. That's the point — and it means the stakes are higher than pure-software tools.

Plugins execute arbitrary code. When plugins are enabled, the agent can create and run Python code on your machine with full server privileges. Review agent-generated plugins before loading them.

Writes affect real hardware. A bad memory write or flash operation can brick a device, wipe calibration data, or trigger unintended behavior. Consider what the agent can reach.

Use tool approval deliberately. When your MCP client prompts you to approve a tool call, consider whether you want to allow it once or always.

This software is provided as-is under the MIT License. You are responsible for what the agent does with your hardware.

License

This project is licensed under the MIT License — see LICENSE for details.