MCP SteamOS Devkit
Manages SteamOS Devkit devices, enabling discovery, pairing, title upload, and remote diagnostics via SSH, ADB, and devkit protocols.
README
MCP SteamOS Devkit
mcp-steamos-devkit is a local MCP server for SteamOS Devkit Client devices. It wraps Valve's current Devkit Client model where possible and includes a native fallback for the core HTTP, SSH, rsync, and remote helper-script flow.
The server is built for Windows-first SteamOS Devkit installs, but the core adapter also works on Linux/macOS when ssh, rsync, and the Devkit helper scripts are available.
What It Exposes
- Device discovery over
_steamos-devkit._tcp.local - Resolve by mDNS name, hostname, or IP
- Device properties and
steamos-get-status --json - Pairing/registration using the Devkit HTTP
/registerflow - SSH key generation and Windows private-key ACL repair
- Sync of
devkit-utilsto~/devkit-utils - Title upload with rsync and Steam shortcut registration
- Title launch through
steam-devkit-rpc run-game - Steam client/session controls
- Logs, controller config dumps, screenshots, GPU traces, RGP captures, RenderDoc replay server control
- Steam Frame native diagnostics: Lepton container/debug-port inventory, Steam/SteamVR/OpenXR log manifests, systemd service listing, journal tails, and bounded OS/Lepton/static-binary inventory for MCP tool discovery
- Steam Frame Lepton ADB helpers: device listing, Wi-Fi/USB connect, forward/reverse, APK install, shell, logcat, bugreport, and Unreal Insights setup
- Steam Frame Unity split APK/OBB helpers: APK package inspection, OBB layout validation/staging, and focused launch diagnostics
- Safety confirmation tokens for destructive or arbitrary remote operations
Quick Start
Clone this repository, then run from the clone directory:
cd C:\path\to\mcp-steamos-devkit
python -m pip install -e .[dev]
mcp-steamos-devkit doctor
mcp-steamos-devkit serve
If the server cannot find the installed SteamOS Devkit Client, set:
$env:STEAMOS_DEVKIT_CLIENT_ROOT = 'E:\SteamLibrary\steamapps\common\SteamOSDevkitClient\windows-client'
Optional source checkout:
$env:STEAMOS_DEVKIT_SOURCE_ROOT = 'C:\path\to\steamos-devkit'
Optional ADB path for Steam Frame Lepton/Android work:
$env:ADB_PATH = "$env:LOCALAPPDATA\Android\Sdk\platform-tools\adb.exe"
Optional SSH password fallback if the device has not accepted the devkit SSH key:
$env:STEAMOS_DEVKIT_SSH_PASSWORD = 'your-device-password'
Optional Android build-tools path for APK metadata inspection:
$env:AAPT_PATH = "$env:LOCALAPPDATA\Android\Sdk\build-tools\<version>\aapt.exe"
Steam Frame ADB
Steam's bundled docs describe ADB as targeting the Lepton Android container on Steam Frame, not the native SteamOS Linux shell. The container is not always running; launch Lepton Development or any Android game first.
Wi-Fi flow:
adb_environment_conflict_doctor(host="frame")
adb_connect_lepton_wifi(host="frame", port=5555)
adb_devices()
adb_logcat(serial="frame:5555")
USB flow:
adb_connect_lepton_usb(local_port=5555, remote_port=5555)
adb_devices()
Useful follow-ups:
adb_install_apk(apk_path="C:\path\game.apk", serial="frame:5555")
adb_bugreport(output_path="C:\path\frame-bugreport.zip", serial="frame:5555")
adb_unreal_insights_setup(tracehost="127.0.0.1", port=1981, serial="frame:5555")
adb_lepton_app_diagnostics(package_name="com.example.game", serial="frame:5556")
adb_shell is intentionally confirmation-gated because it can run arbitrary commands inside the Android container.
Steam Frame Native Diagnostics
These read-only tools inspect the native SteamOS side of Steam Frame:
lepton_containers(target="frame")
lepton_logcat(target="frame", context="steamlaunch-3570175983", lines=300)
lepton_context_inspect(target="frame", context="dev", include_mounts=true)
lepton_debug_targets(target="frame", context="dev")
lepton_mounts(target="frame", context="dev", category="obb")
lepton_apk_info(target="frame", apk_path="~/devkit-game/mygame/game.apk")
lepton_rootfs_overlay_manifest(target="frame")
lepton_debug_plan(target="frame", context="dev", mode="gdb")
lepton_artifacts_manifest(target="frame", context="dev", package_name="com.example.game")
steam_logs_manifest(target="frame", pattern="xrclient", limit=20)
steam_frame_perfcriteria(target="frame")
steam_frame_cef_pages(target="frame")
steam_frame_web_ports(target="frame")
steam_frame_dbus_manager(target="frame")
steam_frame_manager_properties(target="frame", bus="both")
steam_frame_manager_interfaces(target="frame", include_system=true)
steam_frame_openxr_status(target="frame")
lepton_graphics_debug_status(target="frame")
steam_frame_tracking_datasets(target="frame")
sync_tracking_dataset(target="frame", output_folder="C:\path\to\datasets")
local_steamvr_automation_inventory()
steam_frame_automation_inventory(target="frame")
steam_frame_automation_plan(scenario="pose_replay")
steamvr_vrcmd_capability_inventory()
tracking_dataset_analyze(dataset_path="C:\path\to\dataset")
steam_frame_replay_script_template(kind="pose_driver_replay")
deckard_power_status(target="frame")
pidbridge_status(target="frame")
deckard_runtime_environment(target="frame")
native_adbd_status(target="frame")
coredump_list(target="frame", limit=20)
steam_services(target="frame", scope="user", pattern="steamvr")
journalctl_tail(target="frame", unit="steamvr.service", scope="user", lines=200)
steam_frame_dev_inventory(target="frame")
steam_frame_dev_inventory is a bounded discovery pass for future MCP controls. It collects OS build
metadata, Lepton CLI help, Lepton script functions/env references, relevant system/user services,
DBus names, and static strings from known Steam Frame helper binaries. It does not kill, restart,
trace, or mutate Steam/Lepton state.
The Lepton and Steam Frame manager tools above came from inspecting the live Steam Frame runtime, Lepton scripts, DBus introspection, and static binary strings. They are intentionally read-only: they expose container labels, debug-port targets, rootfs overlay files, Deckard runtime files, power state, pidbridge service/socket state, and DBus method/property inventories without starting debug sessions or changing SteamOS Manager state.
The second-pass tools line up with Valve's Steam Frame docs for OpenXR runtime setup, RenderDoc and
Vulkan-layer launch debugging, perfetto/strace artifact handling, and tracking dataset collection.
They report existing state by default; sync_tracking_dataset is a local artifact download helper
for datasets recorded through the headset's SteamVR Developer settings.
Steam Frame Automation
The automation tools focus on repeatable game testing without pretending there is a public controller pose injection command when one has not been found:
local_steamvr_automation_inventory()
steam_frame_automation_inventory(target="frame")
steam_frame_automation_plan(scenario="full")
steam_frame_automation_plan(scenario="launch_regression")
steam_frame_automation_plan(scenario="pose_replay")
steam_frame_replay_script_template(kind="full")
steam_frame_tracking_datasets(target="frame")
sync_tracking_dataset(target="frame", output_folder="C:\path\to\datasets")
tracking_dataset_analyze(dataset_path="C:\path\to\datasets\latest")
steamvr_vrcmd_capability_inventory()
steam_frame_cef_pages(target="frame")
steam_frame_web_ports(target="frame")
adb_logcat(serial="frame:5555", lines=300)
Current reliable lanes are launch/log regression, Lepton ADB input/log capture, Steam/SteamVR CEF UI
inspection, and SteamVR tracking-dataset collection/download. local_steamvr_automation_inventory
also checks the host SteamVR install for the null driver and Frame controller input profiles, which
are useful for PC-side synthetic HMD baselines and binding validation.
steamvr_vrcmd_capability_inventory scans local SteamVR binaries for capture/replay/polling and
driver-simulation strings. Treat those as capability leads, not proven Steam Frame controls, until a
separate tool validates the command behavior. steam_frame_replay_script_template returns JSON
shapes for launch regression, ADB input replay, CEF UI replay, tracking-dataset capture, and future
custom-driver pose replay. tracking_dataset_analyze summarizes downloaded datasets so captures can
be compared across known-good and failing runs.
Controller or HMD pose replay is tracked as a future build path. The docs and live files expose
recorded tracking datasets under ~/.config/openvr/config/cv/xrservice/datasets, but no validated
public Steam Frame CLI for replaying those datasets or injecting arbitrary controller poses has been
found. The practical next implementations are a confirmation-gated CEF replay recorder, curated ADB
input replay, declarative launch smoke tests, and then a custom OpenVR server driver or OpenXR API
layer if scripted pose control is required.
DBus control methods, Lepton debug-server lifecycle, RenderDoc/Vulkan layer injection, tracking dataset packaging, Mesa debug package installation, and coredump debugger backtraces should be implemented as separate confirmation-gated tools because they start capture/debug flows, alter runtime state, or may expose private data.
Steam Frame Unity APK + OBB
Unity split builds can export an APK plus a root-level file named like game.main.obb.
For Android expansion-file loading, the OBB must be visible to the app as:
/sdcard/Android/obb/<package-name>/main.<versionCode>.<package-name>.obb
Steam Frame Lepton maps /sdcard/Android/obb/<package-name> to the title's obb/
directory when that directory exists, so Steam Frame packages should use:
game.apk
obb/main.<versionCode>.<package-name>.obb
Useful flow:
inspect_android_apk(apk_path="C:\path\build\game.apk")
validate_android_split_package(local_dir="C:\path\build")
stage_android_obb_layout(local_dir="C:\path\build")
steampipe_android_release_preflight(local_dir="C:\path\build", app_id="555000", depot_id="555001")
upload_title(target="frame", gameid="mygame", local_dir="C:\path\build", runtime="android")
run_title(target="frame", gameid="mygame")
adb_lepton_app_diagnostics(package_name="com.example.game", serial="frame:5556")
The diagnostics helper reports the app PID/activity state, the OBB symlink target,
SteamAppId/SteamGameId process environment, and logcat highlights for OpenXR,
Steamworks, TMP, exceptions, and common Unity XR failure strings.
steampipe_android_release_preflight is a local read-only checklist for Valve's Android upload
requirements: supported OS, Android depot OS, top-level APK launch executable, obb/ content
layout, package access, Android launch option, and AndroidExternalData cloud-save setup.
MCP Client Config
{
"mcpServers": {
"steamos-devkit": {
"command": "python",
"args": ["-m", "mcp_steamos_devkit", "serve"],
"env": {
"STEAMOS_DEVKIT_CLIENT_ROOT": "E:\\SteamLibrary\\steamapps\\common\\SteamOSDevkitClient\\windows-client"
}
}
}
}
Safety Model
The server treats these operations as confirmation-gated:
- Pairing/registering a device
- Clean uploads that delete remote files
- Local artifact writes such as log syncs and ADB bugreports
- Deleting titles or all titles
- Resetting Steam client state
- Restarting sessions or rebooting
- Clearing Android logcat before capture
- Arbitrary SSH or Steam RPC
- Arbitrary ADB shell commands
Call the tool once without a token to get a requires_confirmation response. Re-run with the returned confirmation_token to execute the same operation.
Notes
The installed Windows client currently ships Python 3.14 bytecode. This package does not require importing that bytecode. It uses Valve-compatible HTTP/SSH/rsync behavior directly, and can also use a source checkout when one is configured.
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。
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。
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。
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。
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。
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。
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。
mcp-server-qdrant
这个仓库展示了如何为向量搜索引擎 Qdrant 创建一个 MCP (Managed Control Plane) 服务器的示例。