SkillAgentSearch skills...

McuBuddy

An AI debugging tool for MCUs and embedded firmware that connects to real hardware through MCP, with support for ST-Link, J-Link, pyOCD, Keil, RTOS, GDB Server, and Flash debugging.

Install / Use

claude mcp add cunjun -- npx -y github:cunjun/McuBuddy

If the server publishes to npm under a different name, use that package instead — check the repo README.

About this skill
🔌

MCP Server

Model Context Protocol server

Quality Score

84/100

Supported Platforms

Claude Code
Claude Desktop

Tags

Our assessment of McuBuddy

McuBuddy scores 84/100 on our quality scale, 249th of 320 Customer Support skills we index.

Its MCP Server is 13 KB long, well organised into 18 sections with 12 code examples: a thorough specification that gives an agent plenty to work with.

It has 10 GitHub stars, so there is little community track record yet; judge it on its content.

Substance
30/30
Structure
20/20
Description
15/15
Adoption
4/20
Freshness
15/15

Maintenance, license and trust

  • The repository was last updated about 3 months ago, so McuBuddy is actively maintained.
  • It is released under the MIT license, a permissive license that allows use, modification and commercial use with attribution.
  • Its trust signals score 97/100, with no cautions. These come from repository metadata, not a code audit — read the skill file before letting an agent act on it.

Safety scan

No issues found

Our scan of the whole file found no instruction hijacking, hidden characters, credential access, data exfiltration or destructive commands. An AI review of the same text found nothing harmful.

AI review by kimi-k2.7-code on 2026-11-07. Automated pattern scan on 2026-11-07. It catches known dangerous patterns, not every risk — read a skill before letting an agent act on it.

McuBuddy compared with similar skills

All 4 of these similar skills score higher than McuBuddy; compare them before choosing.

SkillScoreStarsUpdatedFormat
McuBuddy (this skill)by cunjun84103mo agoMCP Server
Agent-Reachby Panniantong100109.8k1d agoCLAUDE.md
headroomby headroomlabs-ai10079.1ktodayCLAUDE.md
CowAgentby zhayujie10047.8ktodayCLAUDE.md
Scraplingby D4Vinci10099.1k1d agoMCP Server

Frequently asked questions

How do I install McuBuddy?
Run claude mcp add cunjun -- npx -y github:cunjun/McuBuddy. The install tabs above show the steps for each supported agent.
Which AI agents does McuBuddy work with?
It is written for Claude Code and Claude Desktop, as a MCP Server file. Other agents that read the same format can often use it too.
Is McuBuddy safe to use?
Our scan of the whole file found no instruction hijacking, hidden characters, credential access, data exfiltration or destructive commands. An AI review of the same text found nothing harmful. It is MIT-licensed and scores 97/100 on trust signals. Skills are instructions an agent will follow, so read the file before installing it and do not approve commands you do not understand.
Is McuBuddy still maintained?
The repository was last updated about 3 months ago, so McuBuddy is actively maintained.

McuBuddy — AI-Powered MCU and Embedded Firmware Debugging MCP Server

<!-- mcp-name: io.github.cunjun/mcubuddy -->

Python MCP License

Languages: English | 中文

Extend AI from firmware analysis to real MCUs, closing the loop across diagnosis, code changes, build, flashing, and validation in verified environments.

McuBuddy is a Model Context Protocol (MCP) server for MCU board-level debugging. It exposes debug probes, Keil MDK projects, ELF/DWARF symbols, CPU and memory state, SVD peripheral registers, UART/RTT logs, FreeRTOS state, Flash operations, and GDB servers as structured tools that AI assistants can call.

It is designed for firmware development, board bring-up, fault isolation, debugging automation, and AI-assisted validation.

McuBuddy starts with 19 stable tools in the default toolset. Add only the domains a workflow needs with MCUBUDDY_TOOLSETS=probe,diagnose (available domains: probe, diagnose, build_flash, rtos, logs, and experimental). The core profile is the only profile; startup toolset selection is explicit and immutable.

[!IMPORTANT] Automation does not replace engineering responsibility. Humans remain responsible for goals and acceptance criteria, wiring and power safety, high-risk operation approval, code review, and new environment validation. Motors, relays, and other safety-related devices also require recovery plans and independent protection.

Quick links: Quick Start · Project Guide · Tool Reference · Support Matrix

✨ Key Features

  • Real-hardware debugging: Discover and connect to ST-Link, J-Link, CMSIS-DAP, and other probes; control target execution; and inspect registers, memory, breakpoints, and watchpoints.
  • Keil project workflow: Discover .uvprojx / .uvproj files, select a target, invoke Keil MDK through UV4.exe for builds or downloads, and feed the generated AXF/ELF into debugging.
  • Source-level fault diagnosis: Use ELF/DWARF data to resolve addresses to functions, source lines, local variables, and call stacks when investigating HardFaults, startup failures, stack overflows, and memory corruption.
  • Peripheral and RTOS inspection: Decode peripheral registers through CMSIS-SVD and inspect FreeRTOS tasks, task contexts, and stack usage.
  • Logs and runtime observability: Read UART, RTT, and selected J-Link SWO logs, and manage pyOCD/J-Link GDB server lifecycles.
  • Evidence-driven results: Return structured target, state, and validation evidence so AI can continue an investigation instead of guessing code changes from symptoms alone.
  • Actionable hardware boundaries: Distinguish MCU limitations, firmware-inapplicable tools, configuration problems, tool failures, and insufficient evidence, including impact and the next safe check so unsupported paths are not debugged as firmware defects.

🏗️ How It Works

flowchart LR
    AI["AI Client<br/>Codex / Claude Code"] --> MCP["McuBuddy<br/>MCP Server"]
    MCP --> EB["Execution Boundary<br/>Serialized Session"]
    EB --> TOOLS["Debugging Tools<br/>Diagnostics / Symbols / SVD / RTOS / Logs"]
    TOOLS --> KEIL["Keil MDK / UV4.exe<br/>Build / Optional Download"]
    TOOLS --> PROBE["Probe Backends<br/>pyOCD / J-Link / probe-rs"]
    KEIL --> IMAGE["AXF / ELF / HEX / BIN"]
    IMAGE --> TOOLS
    PROBE --> BOARD["Real MCU Board"]

MCP is not a protocol for invoking Keil. The AI calls McuBuddy through MCP; McuBuddy then uses Keil MDK through UV4.exe, pyOCD, J-Link, or another internal backend as required.

🚀 Quick Start

1. Prerequisites

Basic requirements:

  • Python 3.10 or later;
  • a powered MCU development board;
  • a correctly connected ST-Link, J-Link, or CMSIS-DAP probe;
  • the target chip name;
  • preferably, an ELF/AXF image containing debug information.

Keil build and download features require Windows with Keil MDK installed. McuBuddy invokes µVision through UV4.exe, including in Keil MDK v5 installations.

2. Installation

pip install "McuBuddy @ git+https://github.com/cunjun/McuBuddy.git"

This installs McuBuddy once for all local firmware projects. Do not clone or copy the McuBuddy repository into each target project. McuBuddy is a local-only MCP backend: the client starts one stdio process per connection, and McuBuddy does not expose HTTP, SSE, WebSocket, or another MCP network listener.

The target project, Keil installation, ELF/SVD files, probe, and serial port must be directly visible to the machine running McuBuddy. To update, reinstall from the official repository at https://github.com/cunjun/McuBuddy; McuBuddy never checks for, downloads, or installs updates automatically.

Install the optional dependency when using the J-Link Python backend:

pip install "McuBuddy[jlink]"

For development from source:

git clone https://github.com/cunjun/McuBuddy.git
cd McuBuddy
pip install -e ".[dev]"

3. Configure an MCP Client

{
  "mcpServers": {
    "McuBuddy": {
      "command": "McuBuddy",
      "args": []
    }
  }
}

For a Windows source checkout, explicitly configure the virtual-environment Python executable and working directory. See Installation and First Connection, then restart the AI client.

4. Run a First Read-Only Check

After connecting the probe and powering the board, tell the AI:

Use McuBuddy to inspect the current debugging environment, discover connected probes,
and perform a first read-only check of the board without writing Flash.
Before starting, tell me what information is still missing.

The recommended sequence is to check the environment and target first, then configure the probe and read the minimum target state:

doctor()
list_connected_probes()
match_chip_name("py32f030x8")
configure_probe(target="py32f030x8", backend="pyocd")
probe_connect(target="py32f030x8")
read_stopped_context()

probe_connect and read_stopped_context are available in the default core profile. Reading a stable stopped context may halt the target, so it is still execution-changing. If the device must not be halted, instruct the AI to perform only non-intrusive probe and environment checks.

💬 Automated Debugging Example

Use McuBuddy to debug <project path>. The MCU is <exact model>, and the probe is
<ST-Link/J-Link/CMSIS-DAP>. First collect board-level evidence and locate the problem. After
authorization, modify the code, build and flash it, then validate the result on the real board.

For the evidence-first decision order and common scenarios, see Common Debugging Workflows.

🧰 Backends and Hardware Validation

| Path | Current Role | Main Capabilities | | --- | --- | --- | | pyOCD + ST-Link/CMSIS-DAP | Primary backend | Control, memory, Flash, source debugging, RTT, RTOS, and GDB server | | J-Link | Primary backend | Control, memory, Flash, source debugging, native RTT, DWT, and GDB server | | probe-rs sidecar | Extended preview | ARM/RISC-V/Xtensa discovery, configurable core control, registers, memory, hardware breakpoints, Flash, and RTT | | Keil MDK (Windows, via UV4.exe) | Build/download backend | Project discovery, target configuration, build, logs, and optional download; supports MDK v5 installations |

Primary validation coverage includes:

  • STM32L496VETx + ST-Link / pyOCD;
  • STM32F103C8 + J-Link;
  • built-in target preflight profiles for STM32F103ZE and PY32F030X8.

“Implemented in code” does not mean “validated on every board.” Use the Support Matrix and list_validation_records() as the source of truth.

🛡️ Safety Model

McuBuddy provides machine-readable safety classifications through list_tool_safety().

| Category | Examples | Default Requirement | | --- | --- | --- | | Read-only | Target matching, register/memory reads, symbol resolution, logs, diagnostics | No confirmation required | | Execution-changing | halt, resume, reset, continue, stepping | Does not write Flash, but changes execution state | | Runtime-state write | Memory/register writes, breakpoints, watchpoints, SVD field writes | Explicit confirmation | | Persistent destructive operation | Flash erase/program, Keil firmware download | Explicit confirmation | | Host process | Keil build, GDB server start/stop | Starts or stops a local process |

Safety principles:

  1. For an unknown target, match the chip and probe first; do not guess addresses.
  2. Read evidence before halting, resetting, or writing.
  3. Before a Flash operation, confirm the target, scope, image, and recovery method.
  4. For motors, relays, power switches, and other actuators, prefer breakpoints and low-energy tests.
  5. Send actuator commands with uart_send_with_cleanup, then call finish_debug_session before returning a final conclusion. Server shutdown repeats the same idempotent cleanup as a fallback.

🔒 Sessions and Concurrency

  • Operations that share probe, Keil, ELF/SVD, log, and runtime configuration are serialized within the same Session.
  • Different sessions can run concurrently when they control unrelated boards.
  • Stateless queries such as target matching and tool safety information can run alongside session operations.
  • Cancellation cannot forcibly terminate a call that has entered a synchronous SDK. The server waits for the worker thread to finish before releasing the session lock.

This prevents one request from switching backends, disconnecting the probe, or changing shared state while another probe operation is still running.

📦 mcubuddy Skill

The repository includes skills/mcubuddy, which guides Codex and Claude Code to use these tools in an “evidence first, judgment second” sequence instead of treating MCP tools as an unordered command list.

The Skill is an optional workflow enhancement, not a prerequisite for hardware debugging. A correctly installed and configured local McuBuddy MCP server remains fully usable without it.

Installed releases bundle the Skill. Register the persistent Codex integration without cloning the repository:

uv tool install McuBuddy
McuBuddy setup codex --confirm --json

Install for Codex:

python .\skills\mcubuddy\scripts\install_skill.py --target codex --overwrite

Install for Claude Code:

python .\skills\mcubuddy\scripts\install_skill.py --target cc --overwrite

Restart the client or open a new session after installation. For source-checkout recovery, installation registration, and usage boundaries, see Boundaries Between McuBuddy, MCP, and the Skill for details.

⚠️ Current Limitations

  • Keil build and download currently require Windows with Keil MDK and invoke µVision through UV4.exe, including in MDK v5 installations.
  • The probe-rs sidecar covers Flash and RTT but still requires target-specific real-board validation and does not yet have an official binary release.
  • RTOS inspection depends on FreeRTOS symbols and an ELF/AXF that match the target firmware.
  • SVD files are not bundled automatically for every chip and usually come from a CMSIS-Pack or the chip vendor.
  • SWO text capture depends on chip configuration, probe capabilities, pin multiplexing, and board wiring.
  • Device patches and connection strategies remain lightweight mechanisms rather than a complete board plugin system.

📚 Documentation

  • Complete project overview and workflows: [Project Guide](PROJECT_GU

Truncated for display — read the full file on GitHub.

Related Skills

View on GitHub
GitHub Stars10
CategoryCustomer
Updated2mo ago
Forks0

Languages

Python

Trust signals

97/100

From repository metadata: license, adoption, age and documentation. Not a code audit — see the Safety scan above for what the skill file itself contains.

1 info