Keil5 MCP Server
Enables AI coding agents to build, flash, debug, and read feedback from STM32 development projects in Keil MDK, forming a closed loop to automatically fix code.
README
Keil5 MCP Server
English | 中文
A Model Context Protocol (MCP) server that gives deepseek harness a edit code → flash → debug → read feedback → fix code closed loop for STM32 development with Keil MDK.
Instead of manually switching between the IDE, the programmer and the terminal, an agent can:
- Build a Keil project and watch real-time compile progress
- Get structured errors from UV4 logs (file / line / column / code / message)
- Explain error codes with causes and suggested fixes
- Edit source files safely (every edit is auto-backed up)
- Flash firmware via the official UV4 channel or pyOCD
- Debug on hardware through pyOCD: breakpoints, stepping, registers, memory, RTT logs
- Run the official Keil debug channel (UV4
-d+.iniscripts)
Table of Contents
- Features
- Requirements
- Installation
- Quick Start
- MCP Client Configuration
- Tools
- Architecture
- Configuration
- End-to-End Workflow Example
- Safety Rules
- Testing
- Troubleshooting
- Roadmap
- Contributing
- License
Features
- 27 MCP tools registered as
mcp__<serverName>__<tool>(e.g.mcp__keil__build_project) - Real-time build progress: tail-based monitor with percent / current file / phase, capped at 95% until link finishes
- Structured UV4 log parsing: compile errors (
main.c(25:1): error C2065: ...), link errors (L6218E), Program Size, build time - Error-code knowledge base: built-in explanations and fixes for common armcc/armclang codes (C2065, L6218E, L6406E, ...)
- Safe source editing: automatic
.keil-mcp-backups/before every edit, line-range replace, regex search - Official flash path:
UV4 -fuses the project's configured Flash algorithm; pyOCD fallback accepts.axfdirectly - Hardware debug: pyOCD probe control (connect / halt / resume / step / breakpoint / registers / memory / RTT)
- Probe lease: per-probe exclusive access (asyncio lock + file lock) so UV4 and pyOCD never fight over the debug port
- Execution boundary: read-only tools run concurrently; mutating tools serialize on a session lock; cancellation-safe via
asyncio.shield - Works without Keil installed:
keil_doctorreports missing components clearly; the server still starts
Requirements
| Component | Version / Notes |
|---|---|
| Python | 3.10+ (tested on 3.12) |
| Keil MDK | UV4.exe (build -b, flash -f, debug -d) — optional but required for build/flash tools |
| pyOCD | installed automatically via pip; needs a probe driver (ST-Link / J-Link / CMSIS-DAP) |
| Probe | ST-Link V2/V3, J-Link, CMSIS-DAP, Keil ULINKplus |
| Target pack | e.g. pyocd pack install stm32f103c8 or reuse the Keil DFP |
Installation
From PyPI
python -m venv .venv
.venv/Scripts/activate # Windows
# source .venv/bin/activate # Linux / macOS
pip install keil-mcp-server
Package is PyPI-ready (
pyproject.toml+LICENSE+server.jsonincluded). If the package is not yet published, use the source install below.
From source (GitHub)
git clone https://github.com/ZMC1011/dsh-keil-mcp.git
cd ds-keil-mcp
python -m venv .venv
.venv/Scripts/activate # Windows
# source .venv/bin/activate # Linux / macOS
pip install -e ".[dev]"
Verify the install
# Environment self-check (UV4.exe, pyocd, connected probes)
python -m keil_mcp_server --check
# List all registered tools
python -m keil_mcp_server --tools
# Run the unit tests
pytest tests -q
Quick Start
# 1. Start the MCP server (stdio transport — the MCP client will spawn this)
python -m keil_mcp_server
# 2. In your MCP client, call e.g.:
# keil_doctor
# discover_keil_projects { directory: "D:/STM32Projects" }
# configure_keil_project { project: "D:/STM32Projects/app/app.uvprojx" }
# build_project { project: "...", target: "Target 1", stream_progress: true }
# flash_firmware { project: "...", confirm: true }
MCP Client Configuration
DeepSeek Harness (DSH)
Per the official DSH MCP docs: one plugin instance = one MCP server, wired through the official bridge plugin @deepseek-ai/dsh-mcp-client. Add this to your profile's cordis.patch.yml (or cordis.yml):
- insert:
- id: mcp-keil
name: '@deepseek-ai/dsh-mcp-client'
config:
serverName: keil # tools appear as mcp__keil__build_project etc.
transport: stdio
command: D:/000_Environment/mcp-servers/ds-keil-mcp/.venv/Scripts/python.exe
args: ['-m', 'keil_mcp_server']
env:
KEIL_UV4_PATH: D:/002_software/Keil5/UV4/UV4.exe
KEIL_PROJECT_DIR: D:/STM32Projects
# optional: toolCallTimeoutMs: 60000, failOnStartupError: false
Verify with:
dsh web --dump-config | grep -A3 mcp
# or check session logs for mcp__keil__* calls
Note: serverName must match
[A-Za-z0-9_-]{1,32}and be unique among live instances.
Claude Desktop / other stdio MCP clients
Most MCP clients use the mcpServers JSON convention:
{
"mcpServers": {
"keil": {
"command": "D:/000_Environment/mcp-servers/ds-keil-mcp/.venv/Scripts/python.exe",
"args": ["-m", "keil_mcp_server"],
"env": {
"KEIL_UV4_PATH": "D:/002_software/Keil5/UV4/UV4.exe",
"KEIL_PROJECT_DIR": "D:/STM32Projects"
}
}
}
}
For a source checkout without a venv, uv also works:
{
"mcpServers": {
"keil": {
"command": "uv",
"args": ["--directory", "D:/path/to/ds-keil-mcp", "run", "keil_mcp_server"]
}
}
}
Tools
All 27 tools return structured JSON. Destructive operations (flash / erase) require confirm=True.
Build & Errors
| Tool | Description | Key params → Result |
|---|---|---|
build_project |
Compile with UV4 -b (or -r rebuild / -c clean), realtime progress |
project, target?, timeout_seconds?, stream_progress?, clean?, rebuild? → {status, returncode, build_log, errors[], summary, progress?} |
build_progress_status |
Query in-flight build progress | build_id → {status, percent, current_file, phase} |
build_cancel |
Request build cancellation | build_id → {success} |
parse_build_errors |
Parse UV4 log into structured errors | log_path? or log_content? → {errors[], warnings[], summary} |
explain_build_error |
Error code → explanation + causes + fixes | error_code, message?, file?, line? → {explanation, common_causes[], suggested_fixes[]} |
Source Editing
| Tool | Description | Key params → Result |
|---|---|---|
source_read |
Read source with line numbers | file, start_line?, end_line? → {content, total_lines, ...} |
source_edit |
Replace a line range; auto-backup first | file, start_line, end_line, new_content → {success, lines_changed, backup_path} |
source_search |
Search source files (text or regex) | pattern, path?, files?, regex? → {matches[]} |
Official Debug Channel
| Tool | Description | Key params → Result |
|---|---|---|
uv4_debug_session |
Run UV4 -d + generated .ini debug script (headless breakpoint/go/step) |
project, target?, ini_path?, breakpoint?, dump_vars?, timeout_seconds? → {success, returncode, output} |
uv4_debug_dde |
Read session output by id | session_id → {output} |
Project & Environment
| Tool | Description | Key params → Result |
|---|---|---|
keil_doctor |
Environment check: UV4.exe, pyocd, packs, connected probes | — → {uv4_exists, pyocd_installed, probes[], status} |
discover_keil_projects |
Find *.uvprojx under a directory |
directory?, recursive? → {projects[]} |
configure_keil_project |
Parse project: targets, device, pack, groups, source files | project, target? → {targets[], device, pack_id, source_files[]} |
Flash
| Tool | Description | Key params → Result |
|---|---|---|
flash_firmware |
Flash via UV4 -f (preferred) or pyOCD |
project?, image?, backend?, probe_id?, confirm → {success, log} |
erase_flash |
Erase chip flash (pyOCD erase -c) |
confirm, probe_id?, chip? → {success, output} |
verify_flash |
Verify chip against image (pyOCD verify) |
image, probe_id? → {success, output} |
Probe Debugging
| Tool | Description |
|---|---|
probe_connect / probe_disconnect |
Connect / release a pyOCD probe (disconnect frees the port for UV4 -f) |
probe_halt / probe_resume / probe_step |
Core control |
set_breakpoint / continue_target |
Breakpoint by symbol or address, continue |
probe_read_registers |
Read r0-r15, sp, lr, pc, xpsr |
probe_read_memory |
Read memory at address (hex bytes) |
read_rtt_log |
Read SEGGER RTT output (if running) |
Architecture
┌──────────────────────────────────────────────────────────────┐
│ MCP Client (DeepSeek Harness / Claude Desktop / ...) │
│ → tools registered as mcp__keil__* │
└──────────────────────────────┬───────────────────────────────┘
│ stdio (JSON-RPC 2.0)
┌──────────────────────────────▼───────────────────────────────┐
│ keil-mcp-server (Python, FastMCP) │
│ │
│ server.py — tool registration + Execution Boundary │
│ (read-only whitelist → concurrent; │
│ mutating tools → session lock + │
│ asyncio.to_thread + asyncio.shield) │
│ │
│ tools/ — MCP tool layer (27 tools) │
│ │
│ core/ — deliverable layer │
│ uv4_runner.py UV4 -b/-r/-c/-f/-d process runner │
│ build_progress.py realtime log tail monitor │
│ error_parser.py UV4 log → structured errors + KB │
│ source_editor.py read/edit/search + auto-backup │
│ uv4_debug.py UV4 -d + .ini script engine │
│ probe_lease.py per-probe exclusive lease │
│ project_utils.py .uvprojx parser (namespace-tolerant) │
│ │
│ models.py / config.py / config.yaml │
└───────────────┬──────────────────────────────┬───────────────┘
│ │
┌─────────▼─────────┐ ┌─────────▼─────────┐
│ Keil MDK (UV4.exe)│ │ pyOCD + probe │
│ build/flash/debug │ │ ST-Link/J-Link/ │
│ │ │ CMSIS-DAP → chip │
└───────────────────┘ └───────────────────┘
Dependency direction: MCP layer → tools → core → Keil MDK / pyOCD → target chip.
Key design points:
- Execution boundary (inspired by McuBuddy): read-only tools run concurrently; everything else serializes on a per-session
asyncio.Lock, runs in a worker thread (asyncio.to_thread) and is cancellation-protected (asyncio.shield). - Probe lease: UV4
-fand pyOCD cannot share the debug port.ProbeLease(asyncio lock +filelock) serializes access; the flash flow disconnects pyOCD before UV4 takes over. - Realtime progress: a daemon thread tails the UV4 log, counting
compilinglines against the source-file count parsed from.uvprojx(percent capped at 95% until theBuild Time Elapsedmarker). - Malformed-XML tolerance: older Keil projects contain mismatched tags (e.g.
<b498tele498>...</bUseTDR>); the project parser repairs them before parsing.
Configuration
config.yaml (bundled) + environment variable overrides:
keil:
uv4_path: "C:/Keil_v5/UV4/UV4.exe" # or env KEIL_UV4_PATH
default_project_dir: "" # or env KEIL_PROJECT_DIR
build:
build_timeout: 300
stream_progress: true
tail_flush_wait: 3 # seconds to wait for UV4 log tail flush after exit
error:
max_errors: 200
source:
backup_dir: ".keil-mcp-backups"
probe_lease:
lock_dir: ".keil-mcp-locks"
server:
transport: "stdio"
log_level: "INFO"
End-to-End Workflow Example
A typical agent session (tool names shown with DSH prefix mcp__keil__):
1. mcp__keil__keil_doctor # environment + probe OK?
2. mcp__keil__discover_keil_projects # find .uvprojx files
3. mcp__keil__configure_keil_project # parse targets/device/sources
4. mcp__keil__build_project (stream_progress) # compile; on failure:
5. mcp__keil__parse_build_errors # structured errors[]
6. mcp__keil__explain_build_error # causes + fixes
7. mcp__keil__source_edit # fix code (auto-backup)
→ back to 4 until 0 errors
8. mcp__keil__flash_firmware (confirm=true) # UV4 -f → "Verify OK"
9. mcp__keil__probe_connect + set_breakpoint # attach debugger
10. mcp__keil__probe_read_registers / _memory # observe chip state
11. mcp__keil__read_rtt_log # firmware logs
→ if logic bug found: source_edit → rebuild → reflash
Safety Rules
| Level | Operations | Default |
|---|---|---|
| Read-only | chip match, register/memory/symbol reads, logs | no confirmation |
| Execute | halt / resume / step / reset | prompt |
| State write | memory/register writes, breakpoints, watchpoints | confirm |
| Persistent destructive | flash erase / programming | explicit confirm + recovery plan |
| Host process | Keil build, GDB server | prompt |
Principles: gather evidence before acting; identify the target chip first; confirm target / range / image / recovery before flashing.
Testing
pytest tests -q # 11 unit tests: log parsing, source editing, progress, project parsing
Manual smoke tests (in tests/):
python tests/raw_handshake.py # bare JSON-RPC initialize + tools/list over stdio
python tests/func_test.py # end-to-end tool calls through the MCP client SDK
Troubleshooting
| Symptom | Cause / Fix |
|---|---|
Target DLL has been cancelled on flash |
pyOCD still owns the probe. Call probe_disconnect (or let the probe lease handle it) before flash_firmware with the UV4 backend. |
UV4.exe not found |
Set KEIL_UV4_PATH or keil.uv4_path in config; run keil_doctor to confirm. |
No module named keil_mcp_server |
The venv's editable install points at an old path — reinstall from the current checkout: pip install -e . |
No target connected |
Check probe wiring / driver; keil_doctor lists detected probes. |
pyocd pack install needed |
e.g. pyocd pack install stm32f103c8 or point pyOCD at the Keil DFP folder. |
Roadmap
- [ ] Publish to PyPI and register in the MCP registry
- [ ] MCUBUDDY_TOOLSETS-style domain toggles
- [ ] ELF symbol resolution for
set_breakpointby name - [ ] RTOS task awareness (FreeRTOS)
- [ ] GitHub Actions CI for unit tests
- [ ] Linux/macOS support notes (Keil is Windows-only; pyOCD parts are cross-platform)
Contributing
Contributions are welcome! Please open an issue first to discuss changes, then submit a PR.
License
MIT — free to use, modify and distribute with attribution.
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。