AiPC by SYSIK

This tool is for personal automation and testing on your own machine. Do not use it for unauthorized access.

Give the agent "this" — and it gets a real PC.
Screen, mouse, keyboard, your browser, files, terminal, SSH — via the model's native tools.

🌐 Site: aipc.sysik.mom · RU README: Russian version · Skills: SKILL.md · IDE matrix: docs/IDES.md

---

Contents

• What it is
• Why AiPC
• Read this first: risks
• Quick start
• The aipc menu
• 70 tools
• Token efficiency
• MCP connection
• Self-update
• Safety
• What can go wrong
• Troubleshooting
• Project structure
• Build from source
• License

---

What it is

AiPC is a local service + console utility that gives any AI agent full access to your computer over the open MCP (Model Context Protocol) standard.

[ Agent in any IDE ] --MCP/stdio--> [ AiPC-Core: one local service ]
                                            |
        +------------------+----------------+------------------+
        |                  |                |                  |
      Vision            Control          Browser           System/Net
   screenshots,      mouse+keyboard,  your Chrome via    files, terminal,
   UI tree,           apps,            CDP + history,     processes, SSH,
   windows            clipboard        JS eval           search, sysinfo

The agent stops being "text in a chat" and starts working like a human at the PC: it looks at the monitor, clicks, types, opens apps, googles in your browser, runs commands, uses SSH — and verifies every step with a new screenshot.

If a model says "I have no computer access", the built-in system prompt (aipc/server.py:SYSTEM_PROMPT, also served as the aipc_instructions MCP prompt) tells it to use the tools when they are available — or honestly say what's missing.

Why AiPC

Usual agent limits	With AiPC
Blind: no screen	screen_see — screenshot as a native image block, cursor marked
Clicks by guessing	ui_snapshot / ui_find — ready-made x/y coordinates (0–1000)
Typing into the void	focus_type — verified focus + atomic typing; refuses to type blind
Sleeps and hopes	wait_for_window / wait_for_ui_element / wait_for_change
"Did it work?" unknown	screenshot_diff, assert_ui, logs_tail, audit.log
No machine access	Files, terminal, processes, SSH/SFTP, browser, clipboard
Token burn on screenshots	Native image blocks (~65 chars + image vs 55K chars of base64)

---

Read this first: risks

AiPC gives an AI model real control over your PC: it sees the screen, moves the mouse, types, runs terminal commands, reads files, uses SSH. That is the point — and the risk.

• Keep the default ask mode: the server shows you a Yes/No dialog for every dangerous action. The model cannot click "Yes" for you.
• Never switch to auto for tasks involving strangers' files, links or email attachments — prompt injection is real (see docs/THREAT_MODEL.md).
• Know the kill switch: aipc panic or Ctrl+Alt+Shift+K freezes everything.
• This tool is for your own machine (see the note at the top).

---

Quick start

Windows: download AiPC_Win_1.1.1.exe from Releases and double-click it.

Step	What to do
1	On first run it sets everything up itself: asks for admin once (UAC) → copies itself to C:\Program Files\AiPC\ → adds the aipc command to PATH → opens the menu (MCP registration into IDEs is **explicit**: aipc install --ide cursor or menu → Settings → Connect IDE)
2	In your IDE refresh MCP servers (Refresh / restart) and give the agent a plain-language task

macOS / Linux: binaries come from the same Releases page (AiPC_macOS_*, AiPC_Linux_*), or install from source:

npm i -g aipc-sysik   # easiest: downloads the binary for your OS
# or: pip install aipc-sysik
# or: pipx install aipc-sysik
# mirror (needs npm login): npm i -g @sysiknagibator/aipc-sysik
aipc mcp                 # MCP command for your IDE
aipc install --dry-run   # preview IDE changes before applying

The exe above is just the convenient bundle. On macOS grant Accessibility + Screen Recording when asked; on Linux prefer an X11 session (Wayland blocks synthetic input) — details in docs/PLATFORM.md.

Example tasks:

look at my screen, what is this error?
open YouTube and find a review of ...
connect via SSH to 192.168.1.10 and fetch yesterday's log

Works with any tool-capable 2026 model: claude-opus-5-5, gpt-6-astra, gpt-6-sol / luna, gemini-3.8-flash, claude-fable-5-1.

---

The aipc menu

Open a regular cmd and type aipc. One rich.Live frame in an alternate buffer, redrawn only on events — no flicker, no leftover lines. A rounded card holds the logo, slogan and status (core ● running, mode, tools count); below it a frameless numbered list with inline descriptions, groups split by blank lines, the ❯ marker on the selection (danger items turn red when selected, no fills anywhere). One dim footer line with the keys. Controls: W/S, arrows ↑/↓, Home/End, Enter open, digits 1–8 quick select, Q exit, L toggles language (wraps around the edges). Russian layout works by key position. Themes: green/mono/amber (Setup → Theme), NO_COLOR and ASCII fallback respected. Narrow windows hide descriptions automatically; a background check notifies about new releases.

Item	What it does
Launch AiPC-Core	Background server, IDE-style handshake
Stop	Only our own process
Status & self-test	Screen, mouse, terminal, MCP
Diagnostics	Install, PATH, IDE configs
Check updates	Compare with GitHub releases
Setup	Mode ask/auto/read-only, IDE, browser (CDP flag), SSH hosts, language, theme
Service	Reinstall + PATH, reconfigure MCP, kill Core, where-am-I info
Logs	Last 20 lines of audit.log
Exit	—

No menu at hand? Same actions as commands: aipc update, aipc doctor, aipc kill, aipc keys (keyboard diagnostics), aipc status, aipc selftest, aipc setup, aipc install, aipc mcp.

---

70 tools for the model

Vision (18): screen_see, screen_region, screen_burst, window_shot, pixel_color, ui_snapshot (active window by default: 0.69s / 17 nodes vs 1.7s / 200 desktop), ui_find, assert_ui, windows_list, window_focus (verified), window_manage, window_find, get_active_window, wait_for_window, wait_for_ui_element, wait_for_change, screenshot_diff, screen_info.

Mouse & keyboard (19): mouse_move/click/drag/double_click/right_click/middle_click, mouse_position (drag with ctrl/shift/alt), scroll, type_text (auto-chunked, Cyrillic-safe), press_key, key_down/up, clipboard_set/get (+images), sleep, open_app, focus_type (verified focus + atomic type — the only correct way to type).

Browser (6): browser_tabs, browser_goto, browser_eval (JS via CDP — more reliable than coordinate clicks), browser_active_tab, browser_close_tab, browser_history_search (read-only copy of the History DB).

Files & system (13): fs_list/read/write/find/stat/mkdir/delete/move (atomic writes + .bak, binary guard, pagination, deny-lists), run_cmd (split stdout/stderr, Russian OEM decoding, deny-list), process_list/find, wait_for_process, download_file (streamed, no size cap).

Network & machine (7): web_search_pc, ssh_exec, ssh_sftp_get/put (200 MB+ friendly streaming), sys_info (battery/memory/CPU/disks), net_check, env_get (refuses secret-like names).

Human & debug (4): notify_user (popup, non-blocking), ask_user (Yes/No, blocks for a human decision), aipc_status, logs_tail.

Audio (1): audio_listen (loopback — system sound, or mic; 1-30 s, WAV, native audio block or saved file).

Video (2): video_info (duration/size/fps/codec), video_frames (1-12 frames as image blocks — watch a clip frame by frame).

Agent loop: see (screen_see) → do → re-see to verify. Errors are structural: {"ok": false, "reason": "not_found|timeout|denied|..."}. Every call runs exactly once, serialized; every call lands in ~/.aipc/audit.log (2 MB rotation).

Token efficiency

• Screenshots travel as native MCP image blocks: ~65 chars of JSON + image instead of ~55K chars of base64 text (~10x cheaper per screenshot). raw=true returns legacy base64 if a client can't render images.
• UI nodes are compact: {"t","n","x","y"} instead of long keys.
• ui_snapshot defaults to the active window; role/name filters cut tokens further.
• Prefer search over listing: ui_find / window_find / fs_find / process_find.
• All tool descriptions fit in one short line.
• Profiles (aipc mcp --profile ...): minimal (12 tools: see→do→verify loop), browser (20 tools: + web), full (70, default). Measured tool-schema size: full ~12.9K chars, browser ~3.8K, minimal ~2.5K (−81% vs full).

---

MCP connection

AiPC registers itself on every menu launch (19 configs across 14 IDE families). Manual reference format:

{
  "mcpServers": {
    "aipc": {
      "command": "C:\\Program Files\\AiPC\\AiPC_Win_1.1.1.exe",
      "args": ["mcp"]
    }
  }
}

Full 40-IDE matrix (auto / preset / bridge): docs/IDES.md. Ready-made files: mcp_presets/. Skill-style install: SKILL.md.

Cloud IDEs without localhost access (Claude.ai, Manus, Devin, Bolt, v0, Lovable, Replit…) can't reach 127.0.0.1 — see the "Bridge" section in docs/IDES.md.

Self-update

Menu → "Check updates" (or aipc update): compares with the latest GitHub release, shows the changelog, and only with your explicit Yes downloads and reinstalls itself (UAC). Version is always visible in the menu title; aipc doctor flags a stale install in Program Files.

---

Safety

Mode	File writes/deletes	Commands/SSH	Typing in password managers	Who decides
ask (default)	popup to YOU	popup to YOU	popup to YOU	human, every time
auto	allowed	allowed	allowed	model (except after untrusted input — taint-guard asks)
read-only	blocked server-side	blocked server-side	blocked server-side	nobody

• Modes: ask (default — the server pops a Yes/No dialog to YOU for every file-write/delete, command, SSH, download, browser-JS call; the model cannot bypass it; silence for 120 s = denied), auto (no dialogs), read-only (all mutating tools blocked server-side, not just by prompt).
• Risk classes: every tool is tagged read/interact/mutate/exec/network (policy.TOOL_RISK). read-only passes only read; ask confirms mutate/exec/network; typing/clicking inside password-manager, crypto-wallet or banking windows also asks. The dialog offers time choices — once, 10 min, 1 hour (safety.allow_presets) — never forever.
• Honest note: deny-lists only reduce risk, they are not a security boundary (obfuscation always evolves). The real boundary is modes + server-side confirmations. For strict setups use run_cmd_policy: allowlist — only listed command prefixes run, everything else asks.
• Admin rights needed once — copying into Program Files + writing PATH.
• Deny-lists for commands (mimikatz, ransomware patterns, encoded PowerShell, quotes/^/env-var/chain-obfuscation normalized) and paths (browser secrets, private keys, hives, System32) live in ~/.aipc/config.yaml and auto-upgrade (safety.version) without wiping your additions. Paths inside run_cmd (type ...\Login Data) are checked too.
• Typing goes only into a verified foreground window (focus_type).
• Secrets guard: env_get refuses *KEY/*TOKEN/*SECRET/*PASSWORD* names.
• Emergency stop: menu → Stop, or aipc kill (verifies the PID is ours first).

---

What can go wrong

• Prompt injection: a web page, document or chat message tells the agent to delete/send/run something. Mitigations: untrusted markers, taint-guard confirmations, ask by default. Still — read the confirmation dialogs.
• Wrong window, wrong click: the agent acts on what it sees; a changed layout means a wrong button. focus_type refuses to type blind, but stay nearby for destructive tasks.
• Update without signature: aipc update downloads an exe checked only for the MZ header. Don't update over hostile networks.
• Secrets in context: the agent may paste a token it read into chat. audit.log, env_get and fs_read mask known patterns; everything else is your eyes on the screen.

Full version: docs/THREAT_MODEL.md.

---

Troubleshooting

Symptom	Fix
Model says "no PC access"	Reconnect MCP (Refresh / restart IDE), remind: "you have aipc.* tools, start with screen_see"
Menu doesn't respond to keys	Run aipc keys, press W/S/arrows/Enter/Esc — paste the 5 lines to us
aipc command unknown	Open a new cmd (PATH applies to new shells); reinstall via dist\aipc.exe
Right wall of the menu drifts	Update: frames render only via fixed-width rich.Panel since 1.0.4.4
Chrome tabs invisible	Menu → Setup → Browser (adds --remote-debugging-port=9222), restart Chrome
Old version in Program Files	aipc doctor tells you; run fresh dist\AiPC_Win_1.1.1.exe → UAC → updated
MCP not in IDE	Run aipc install --list-ides, then aipc install --ide cursor (or --all); preview with --dry-run
Remove AiPC	aipc uninstall (removes our MCP entries, PATH, Program Files copy; --yes for scripts)
GitHub page/API 404s or rate-limits	Wait a minute and retry; check your token scopes and proxy; if the outage persists, open an issue with the timestamp

---

Project structure

AiPC/
  aipc/            # server.py (MCP) · vision/control/os_ops/browser/net/sysinfo
                   # notify · installer (self-install + 19 IDE configs)
                   # maintenance (update/doctor/kill) · menu · policy/audit
  mcp_presets/     # ready-made MCP configs (16 files)
  assets/          # logo (AiPC_Logo.svg/png) + app icons: AiPC.ico (Windows),
                   # AiPC.icns (macOS) — visible on light and dark themes;
                   # AiPC_promo*.mp4 — video overviews (EN/RU) for README
  tools/           # build_exe.bat · build_icon.py · sign.bat · version_info.txt
  tests/           # pytest: policy, installer merge, server (19 tests)
  docs/            # IDES.md (40 IDEs) · SIGNING.md (SmartScreen)
  SKILL.md         # one-file skill install · CHANGELOG.md · LICENSE (MIT)

---

Build from source

python -m pip install -r requirements.txt
python -m pytest tests -q
python -m aipc selftest
tools\build_exe.bat

Output: dist\AiPC_Win_1.1.1.exe (~36 MB, icon + publisher version-info) and dist\AiPC-Setup.exe. Signing away the blue SmartScreen: see docs/SIGNING.md (needs a code-signing certificate: Certum Open Source ~€25/yr is the cheapest start).

---

License

MIT. Author — SYSIK. See LICENSE.