Human Input MCP
Windows x64 Human Input MCP for safe, observable mouse/keyboard control of explicit HWNDs, plus background reading and timers.
Query capabilities and mode; enable/disable human input; observe target identity, focus, foreground, geometry, DPI, and cursor.
Preview bounded input plans without injecting input; update targets; stop, pause, and resume with safety checks.
Move, click, drag, and perform guarded safe-area activation using physical virtual-screen coordinates, with cursor-inside-client and state checks.
Type printable text with WPM/seed/typo settings, send allowlisted keys/chords, and explicitly copy/cut/paste without inspecting clipboard contents.
Scroll vertically or horizontally in Windows wheel units.
Start, monitor, and cancel background reading traversal over client regions using pointer movement only.
Read operation status and bounded events; acknowledge event receipt; manage external-observation timers.
Cannot capture screens, OCR/read page content, focus or rearrange windows, inspect clipboard contents, detect chat/other MCP calls, or verify remote delivery; callers must handle paused/uncertain states and provide fresh HWND/coordinates/safe areas.
Click on "Deploy Server".
Wait a few minutes for the server to deploy. Once ready, it will show a "Started" state.
In the chat, type
@followed by the MCP server name and your instructions, e.g., "@Human Input MCP enable human mode, activate Notepad, and type a short greeting for me"
That's it! The server will respond to your query, and you can continue using it as needed.
Here is a step-by-step guide with screenshots.
Human Input MCP 0.2.0
Windows x64 的獨立輸入 MCP。Node.js 提供 30 個工具;Rust worker 透過 Windows API 檢查視窗並執行輸入。官方 Touch MCP 保留原版,獨立安裝、啟動及更新。每個新程序的 human mode 預設關閉。
安裝與移機
一般使用請解壓縮完整可攜版,執行完整可攜版根目錄的 launch.ps1。套件內附 Node.js 24.22.1、預編譯 static CRT Human worker、監督程式、Node 相依套件,以及官方 Touch 所需的 .NET 10.0.13。正常使用不需要系統 Node、Rust 編譯器或 Python。官方原生模組仍需 Microsoft Visual C++ v14 x64 Runtime;啟動器會檢查並提供官方安裝資訊。
.\launch.ps1 version human
.\launch.ps1 version touch
.\launch.ps1 verify human --boot
.\launch.ps1 verify touch --boot
.\launch.ps1 config human
.\launch.ps1 config touch
.\launch.ps1 launch human
# 官方 Touch 使用另一個 MCP 程序:
.\launch.ps1 launch touch移到新電腦後,重新執行 config human 與 config touch,取得當地路徑的設定片段。設定輸出不會改寫現用 MCP 設定;由使用者在需要時自行登錄。完整命令與來源驗證見 可攜版指南。
Related MCP server: windows-computer-use-mcp
操作順序
外部觀察工具先確認目標用途、內容、HWND 與實體螢幕座標。
查詢
human_capabilities {},再明確呼叫human_mode {"enabled":true}。以
human_observe {"hwnd":"…"}取得最新視窗身分、前景、焦點、幾何、DPI 及游標狀態。這不代表已確認畫面內容。每次只呼叫一個明確語意操作,再由接收端或外部觀察確認結果。支援移動、點擊、拖曳、逐字輸入、捲動、允許的按鍵,以及獨立的複製/剪下/貼上。
結束時呼叫
human_stop {},再以human_mode {"enabled":false}關閉模式。
完整的「Notepad 點按鈕 → 打一句話 → Enter → 掃讀區域」與復原步驟見 Agent 操作流程。HWND 查找由外部觀察工具負責,產品不新增視窗探索工具。
所有點、safe_area、掃讀矩形及 observe client 矩形均採實體虛擬螢幕絕對像素,可為負數;不是 client 相對座標、DIP 或縮放截圖像素。例如 observe 的 client 原點為 (100,100),按鈕在 client 內偏移 (200,200),應傳 (300,300)。移動、點擊、拖曳與掃讀的路徑全程留在目標 client 內,游標必須已在 client 內,否則回傳 cursor_outside_client;只有 human_activate 能從外面進入。輸入前及執行中檢查身分、前景、焦點、遮擋、幾何、DPI、游標與鍵盤狀態。沒有焦點切換 API 或視窗重排。human_activate 是獨立的安全區域點擊,必須提供最新 safe_area 及 safe_area_confirmed:true;確認區域不會編輯、導航或提交。沒有可確認的 client 安全區域時,請操作者將視窗置前並把游標移入 client,再重新觀察;標題列不能充當 safe_area。capture_conflict 表示滑鼠 capture 衝突,須先由操作者解除並重新觀察。點擊後重新觀察,再決定後續操作。
指標 duration_ms 為 100–15000 ms,僅表示移動/接近時間;點擊、拖曳與 activation 的按住/釋放時間另加,總計畫上限為 15200 ms。預算錯誤以 invalid_plan 說明 required_ms 與 limit_ms。
human_type 預設 41 WPM、typo_rate:0;刻意錯字須明確啟用。text schema 的 maxLength 為 1000,規劃器另限 1000 codepoints 與 45000 ms;預設速度約可輸入 150 個 ASCII 字元,較長文字須分次呼叫。預算錯誤另提供該 WPM 的建議字數。每次 typing 回應與狀態皆包含有效 seed。逐字輸入不暗中貼上,也不接受換行或 Tab;需要 Enter 時使用獨立 human_key。按鍵先去除頭尾空白、將 CamelCase/PascalCase 轉 snake_case 並小寫化,例如 Enter → enter、PageDown/PAGE_DOWN → page_down;完整 canonical 清單及 ctrl/shift 組合見 工具契約。一般按鍵拒絕 Ctrl+C/X/V,剪貼簿操作必須使用 human_copy、human_cut、human_paste。產品只送出 Windows 快捷鍵,不讀取或保存剪貼簿內容。剪下與貼上可能更改接收端內容。
所有同步輸入都可能回傳 state:"paused" 而尚未完成;必須檢查 state,才能決定是否進行下一步。錯誤與暫停結果的 recovery 提供 IME、焦點/幾何/身分改變、游標、capture、權限、預算與恢復拒絕等修復方向。ime_open 若尚無效果送出,可 stop、重新觀察,再以 force_ime_off:true 建立新 typing;或由操作者關閉 IME、恢復原始鍵盤 layout 與 Caps Lock 後,以目前 revision 明確 continue,不能在舊工作直接改選項。初始編譯與恢復皆用 fresh-thread Caps Lock 快照比較;相符才繼續其餘檢查,不符回覆 keyboard_state_changed 並保留乾淨暫停狀態。無法還原或核對不明效果時,stop、重觀察並重新規劃。
背景掃讀與工作控制
reading_start 接受有順序、目的、方向與時間預算的區域,立即回傳 operation_id,背景只做純游標移動與停留,總長最多 60 秒。格式與界線見 工具契約。狀態、觀察、事件及倒數查詢不會取消掃讀;新輸入、目標/模式變更或停止先經過取消屏障。另一個掃讀需要明確 replace:true。
游標異常立即暫停,最多觀察 3000 ms。短暫小幅且穩定的移動只有在重新檢查安全後才可自動接續;持續大範圍移動成立後回報 pointer_takeover 並維持暫停。數值門檻可設定,目前是工程預設值,尚未宣稱實測校正或能辨識操作者。
human_pause 回報清理屏障與進度;暫停時拒絕所有新輸入。手動恢復須提供目前 operation_id、精確 pause_revision、external_checked:true 與明確 resolution。已送出的效果不會重播;不明效果須依外部證據逐項解決,不能盲目 continue。停止、斷線、模式關閉及程序重啟後不自動恢復。詳見 原生協定。
journal 附加 unit_kind、char 單元的 text 及零起算 source_index,優先保留目前不明單元,再補最近已送出 frames。Typing 的 uncertain_units_detail 另標示 intended_char、typo_char 或 typo_correction_backspace;無法對照時標示 unknown,須停止、重新觀察與規劃,不能猜測效果。暫停期間焦點/幾何改變須外部復原;此時不能啟動 human_activate 修復。
無即時工作時,human_pause {} 仍會鎖住新輸入的接納;唯讀查詢與模式/目標變更不解除這個 idle pause latch,須明確 human_resume {} 或 human_stop {}。
completed 表示執行器完成,inputInjected 表示已送出原生輸入;deliveryVerified:false 不證明接收端結果或遠端送達。取消確認後不再送出新輸入,已送入 Windows 的事件無法撤銷。
事件、倒數與限制
human_events 保留有界事件,human_event_ack 只記錄 MCP 呼叫者確認,不會恢復工作。可選的診斷通知與測試接收不代表 Codex Agent 正在思考時收到即時中斷;目前 agent_push_supported:false,實際 Agent 即時送達仍不支援/未驗證。
分頁以 next_seq 接續:它是本次最後回傳事件的序號,沒有事件時保留 after_seq。latest_seq 是最新保留事件的序號,不能取代分頁游標。主動工作收到 terminal 事件後會抓取最終狀態、清除 active 並停止輪詢。
倒數使用 timer_attach → timer_prepare → 外部讀取 → timer_calibrate → timer_status。單調時鐘估計與外部觀察值分開,納入讀取/傳送延遲及未知顯示取整。一般 60 秒、最後 30 秒內 10 秒超齡;換頁、暫停或恢復後須重新校正。估計歸零不觸發輸入。
timer_invalidate.reason 若提供,長度須為 1–512 字元。
Human Input 不擷取畫面、OCR、讀取頁面內容、安裝 hooks,或偵測其他 MCP 呼叫/聊天訊息;cancel_on_user_message:false、page_change_detection:false。多個 Human worker 共用輸入鎖,但無法約束官方 Touch 或其他軟體;改用其他輸入工具前先停止 Human 工作。權限、UIPI、IME、快速實體輸入及外部強制終止仍有平台限制。沒有所有 Windows/鍵盤/DPI 組合或全新 VM 相容性保證。
更新、回退與 0.1 遷移
先停止 Human 工作並正常關閉相關 MCP 程序,再明確執行更新。更新前校驗可信 SHA256、封存路徑及完整 manifest,使用 staging 與唯讀探測;整棵探測程序樹確認退出後才原子切版。失敗保留原選版,成功保留前版,回退離線重驗。兩個元件獨立選版與鎖定。
.\launch.ps1 update human --bundle 'D:\下載\human-input-0.2.0.zip' --sha256 '可信的64字元SHA256'
.\launch.ps1 update human --url 'https://使用者明確指定的來源/release.zip' --sha256 '可信的64字元SHA256'
.\launch.ps1 update touch --version 2.3.2
.\launch.ps1 rollback human
.\launch.ps1 rollback touch沒有自動更新或預設 Human 發行站。0.1 的 Node 與 protocol 1 worker 不可混用新版。改用完整 0.2 可攜版,正常停止舊程序,產生新路徑設定,再由使用者切換;保留舊安裝可供回退。錯字預設值改為零,按鍵、剪貼簿、安全啟動及暫停/恢復採新版契約,呼叫者須處理完整狀態與進度,不能把舊停止結果當成可恢復狀態。
開發與驗證
以下命令屬於原始碼 checkout 的開發/驗收流程;完整可攜版根目錄不保證包含測試工具或工作目錄。開發才需要 Rust/MSVC。準備已驗證的內附 Node 後:
& './work/prerequisites/node-v24.22.1-win-x64/node.exe' --test 'tests/*.test.mjs'
cargo test --locked --manifest-path native/Cargo.toml
powershell -NoProfile -ExecutionPolicy Bypass -File scripts/build.ps1 -NoCopy
$env:HUMAN_INPUT_WORKER=(Resolve-Path 'native/target/portable/x86_64-pc-windows-msvc/release/human-input-worker.exe').Path
& './work/prerequisites/node-v24.22.1-win-x64/node.exe' scripts/integration.mjs可見驗收 node scripts/live-verify.mjs 會開兩個專用 WinForms 視窗,自動對它們送出真實輸入並比對結果,約 40 秒,執行期間不要碰滑鼠鍵盤。選項與判定方式見 可見驗收操作,最新結果見 驗證紀錄。
第三方來源與授權見 THIRD-PARTY-NOTICES.md。每次封存的來源版本、雜湊、搬移與啟動結果,以該次外部交付報告及 build-provenance/manifest 為準。目前自動可見驗收 18/18 通過(含剪貼簿、多區域掃讀軌跡,以及腳本化的游標小幅偏移自動接續與大幅接管);真人移動滑鼠的冒煙測試尚未以新版指示重跑,詳見驗證紀錄。
Available Tools
30 toolshuman_activateCDestructive
Distinct guarded safe client click. Caller must freshly confirm the safe area does not edit, navigate or submit. Reobserve before any following input.
Inputs: x (required), y (required), hwnd (required), safe_area (required), +6 more.
human_capabilitiesBRead-only
Report human mode and truthful limits. This service cannot see chat messages, page contents, or calls to other input tools.
Inputs: no named parameters.
human_copyCDestructive
Explicit Ctrl+C using the system clipboard; contents never inspected.
Inputs: hwnd (required), pause_policy, threshold_px.
human_cutCDestructive
Explicit Ctrl+X; may modify content. No clipboard content inspection.
Inputs: hwnd (required), pause_policy, threshold_px.
human_event_ackCDestructiveIdempotent
Acknowledge this MCP caller receipt using the event token. Never resumes input or proves model receipt.
Inputs: event_id (required), receipt_token (required), session_epoch (required), pause_generation (required).
human_eventsCRead-only
Read bounded event history. Discovery is delayed; agent push is unsupported.
Inputs: limit, after_seq, pending_only.
human_input_planARead-only
Preview a bounded plan without input or cancellation. Typing previews are pure and need no hwnd; pointer previews observe explicit hwnd for client bounds. Page freshness is the caller responsibility.
human_keyCDestructive
One allowlisted editing/navigation chord. Enter may submit. Clipboard actions require dedicated tools.
Inputs: key (required), hwnd (required), modifiers, pause_policy, +1 more.
human_modeADestructive
Query current mode with empty arguments, or explicitly enable/disable human input. Every process starts disabled; a mode change cancels current work before returning.
Inputs: enabled.
human_mouse_clickADestructive
Move to a client point and pulse the left button once; waits for completion. The cursor must already be inside the target client. Caller must freshly establish the intended UI target.
Inputs: x (required), y (required), hwnd (required), jitter_px, +4 more.
human_mouse_dragADestructive
Drag with the left button from the current observed cursor to a client point; waits for completion. Does not pre-position the cursor.
Inputs: x (required), y (required), hwnd (required), jitter_px, +4 more.
human_mouse_moveADestructive
Move from the observed cursor through a bounded human-style path; waits for completion. The cursor must already be inside the target client (use human_activate to enter). Cancels background reading first. No auto focus.
Inputs: x (required), y (required), hwnd (required), jitter_px, +4 more.
human_observeARead-only
Read native target identity, client geometry, focus, foreground, DPI and cursor. Does not capture images, focus a window, or cancel reading.
Inputs: hwnd (required).
human_pasteCDestructive
Explicit Ctrl+V; may modify content. No clipboard content inspection.
Inputs: hwnd (required), pause_policy, threshold_px.
human_pauseDDestructive
Pause with cleanup barrier, or latch admission while idle.
Inputs: operation_id.
human_resumeCDestructive
Explicit checked continuation. Fresh native observation follows pause cleanup. Empty arguments only clear idle pause; uncertain effects require exact external outcomes.
Inputs: resolution, operation_id, pause_revision, external_checked.
human_scrollCDestructive
One explicit wheel operation in Windows units, multiples of 120 up to 1200. No pixel/line inference.
Inputs: hwnd (required), delta (required), axis, pause_policy, +1 more.
human_statusBRead-only
Readonly operation status and pending safety events. Does not resume or cancel.
Inputs: operation_id.
human_stopADestructiveIdempotent
Cancel current or pending input and wait until no future injections are possible. Does not return the cursor or release user-owned input.
Inputs: reason.
human_target_updateCDestructive
Cancel current work with a barrier, then observe and replace the selected target. Never focuses or moves a window.
Inputs: hwnd (required).
human_typeADestructive
Type printable text using bounded character pulses and optional corrections; no paste, Enter, or arbitrary chords. IME change needs force_ime_off=true; remote delivery is not verified.
Inputs: hwnd (required), text (required), wpm, seed, +4 more.
reading_cancelBDestructiveIdempotent
Cancel reading with a no-future-injection barrier. An omitted operation_id selects the active reading operation.
Inputs: reason, operation_id.
reading_startADestructive
Begin autonomous pure pointer traversal of explicit client regions and return promptly. The cursor must already be inside the target client. No buttons or keys. New input, stop, target or mode changes cancel it; chat messages cannot.
Inputs: hwnd (required), segments (required), replace, pause_policy, +1 more.
reading_statusBRead-only
Read reading operation status without cancelling it. deliveryVerified=false; no page or reading-comprehension detection.
Inputs: operation_id.
timer_attachCDestructive
Attach an external-observation timer to an explicit target and section; never reads screen pixels or cancels input.
Inputs: target_id (required), section_id (required), timer_id.
timer_calibrateBDestructive
Anchor a timer using a prepared ticket and caller-observed seconds. Observed expiry requires explicit expired_observed state.
Inputs: timer_id (required), ticket_id (required), observed_seconds (required), state.
timer_detachCDestructive
Detach one external-observation timer without touching input.
Inputs: timer_id (required).
timer_invalidateBDestructive
Invalidate a timer estimate after known external changes; does not inspect or mutate a page.
Inputs: timer_id (required), reason.
timer_prepareCDestructive
Prepare a bounded observation ticket before the caller reads a timer through an external source.
Inputs: timer_id (required).
timer_statusBRead-only
Return the observed timer anchor and an uncertain estimate separately. Querying never cancels input.
Inputs: timer_id (required).
Tool Schema Changelog
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
Loading tool schema changelog…
TDQS
Scored across 30 tools
The tools are grouped into human input, reading, and timer domains, and descriptions clarify many boundaries. However, several similar safety/status/cancel tools exist (human_stop, human_pause, human_mode, human_status, reading_cancel, timer_invalidate/detach), and input tools like human_activate, human_mouse_click, human_key, and human_type could still be confused without careful reading.
Names use consistent snake_case and clear domain prefixes: human_, reading_, and timer_. Most follow prefix_action or prefix_noun patterns, with minor deviations such as human_key and human_capabilities.
Thirty tools is heavy for this server, even accounting for the safety, reading, and timer subdomains. The surface feels over-granular, with many specialized status, cancellation, and timer tools that could be consolidated.
The surface covers the core human-input lifecycle: observation, targeting, pointer and keyboard actions, pause/resume/stop, event history, reading traversal, and timer coordination. Some gaps remain, such as no explicit right-click/double-click or window-focus tool, but those may be intentionally out of scope.
Maintenance
Related MCP Connectors
Eyes and hands on real Windows PCs — observe, click, type via Glasswarp API.
MCPs that do tasks on any website. Search flights, book, shop, check accounts and fill in forms.
Securely control computers you explicitly pair through files, terminals, processes, screenshots, desktop UI/input, clipboard, browser automation, diagnostics, and document tools.
Build and run AI workflows, agents, data tables and 1000+ API integrations from any MCP client.
Related MCP Servers
- FlicenseNot gradedqualityDmaintenanceEnables automation of native Windows desktop applications through screen capture, mouse/keyboard control, and waiting for UI changes, exposing them as MCP tools.1-
- AlicenseNot gradedqualityDmaintenanceProvides desktop automation tools for Windows, enabling mouse and keyboard control, screenshots, clipboard management, and application interaction via the MCP protocol.151 npmMIT
- AlicenseNot gradedqualityCmaintenanceEnables AI clients on Windows to control the local mouse, keyboard, and screen understanding via MCP stdio, allowing automated workflows like viewing the screen, locating elements, clicking, typing, and verifying results.1Apache 2.0
- AlicenseNot gradedqualityBmaintenanceEnables MCP agents to drive Windows applications by targeting individual windows—clicking, typing, scrolling, dragging, and reading accessibility trees or screenshots—without stealing the user's focus, mouse, or keyboard. Runs entirely locally alongside the user, with per-app approval and no cloud account or usage limits.MIT