Skip to main content

​
AI Use Computer with Kernel CLI & MCP — Free & Freedom for All AIs

In 2026, only a few AIs can “use a computer.” Claude charges $20-200/mo. OpenAI requires a subscription. Both lock you to their model. Both use slow screenshots. Both decide your safety boundary for you. Kernel CLI gives every AI computer use — for free, through MCP.

​
One Config. Every AI Gets Computer Use.

{
  "mcpServers": {
    "kernel-cli": {
      "command": "kernel-cli",
      "args": ["--mcp"]
    }
  }
}
Add this to any AI that supports MCP. Done. 
Claude Desktop   → can now press buttons, read screens, control apps
Claude Code      → can now merge PRs in Chrome, notify in Slack
Codex CLI        → can now check CI status, manage windows
Gemini CLI       → can now query browser history, launch apps
Cursor           → can now read any app's UI structure
Any MCP AI       → full OS access
No subscription. No vendor API. No VLM cost. Local binary. Instant.

​
What Every AI Gets

ToolWhat it doesExample
ax_treeRead UI structure of any appSee every button in Chrome
ax_pressPress any button by name”Merge pull request”
type_textType text into any fieldWrite a Slack message
keyKeyboard shortcutsCmd+S, Cmd+Enter, Ctrl+C
windowsList all open windowsFind which app is focused
snapshotFull system snapshotWindows + AX tree + clipboard
appLaunch, quit, switch appsOpen Safari, close TextEdit
processRunning processes + CPU/memoryWhat’s using resources
querySearch browser history, Notes, Photos”What did I search yesterday?”
captureScreenshot (when visual is needed)Last resort, not first
scriptRun AppleScript/JXAComplex multi-app automation
scrollScroll any elementNavigate long pages
windowResize, move, focus windowsArrange workspace
searchSpotlight searchFind any file
One MCP server. 14+ tools. Every AI.

​
Crazy Fast. No Screenshots.

Every other “computer use” works like this: 
Screenshot approach (Claude CU, OpenAI Operator):
  📸 Take screenshot                    500ms
  🧠 Send to Vision Model (VLM)        2000ms
  🤔 VLM guesses: "button at (450,320)" 1000ms
  🖱️ Click at coordinates               100ms
  ❌ Might miss. Retry.                 +3600ms

  Total: ~3.6 seconds. Probabilistic. Costs money per screenshot.
Kernel CLI: 
AX tree approach:
  🌳 Read accessibility tree             50ms
  🔍 Find "Merge pull request"            1ms
  ✅ Press via AX API                     10ms

  Total: 61ms. Deterministic. Free.
Crazy fast. Never misses. Zero cost. The accessibility tree is how screen readers navigate computers. It’s the OS-native API for understanding UI. Kernel CLI speaks the language the OS already speaks.

​
Real: Claude Desktop Merges a PR

Claude Desktop has no terminal. It can’t run shell commands. But with Kernel CLI MCP, it can control Chrome: 
Step 1: Read Chrome's UI
  → ax_tree(app="Chrome")
  → AXButton "Merge pull request" ✓
  → AXStaticText "All checks passed" ✓

Step 2: Press the button
  → ax_press(title="Merge pull request")
  → PR merged.

Step 3: Switch to Slack
  → app(action="switch", name="Slack")
  → ax_tree(app="Slack")
  → AXTextField "Message #engineering"

Step 4: Type notification
  → type_text("Feature shipped. PR #142 merged.")
  → key(combo="enter")
  → Message sent.
Four MCP calls. No screenshot. No VLM. No coordinates. Claude Desktop just merged a PR and notified the team.

​
Real: Codex Checks CI and Deploys

Codex CLI runs in a terminal but can’t see Chrome. With Kernel CLI MCP, it can: 
Step 1: Check CI
  → ax_tree(app="Chrome", depth=3)
  → AXStaticText "All 47 checks passed"

Step 2: Read deploy button state
  → AXButton "Deploy to production" [enabled]

Step 3: Deploy
  → ax_press(title="Deploy to production")

Step 4: Verify
  → ax_tree(app="Chrome")
  → AXStaticText "Deployment in progress..."
Codex never left the terminal. But it saw Chrome, read CI status, and pressed Deploy — through MCP.

​
Real: Gemini Researches Across Apps

Step 1: What tabs are open?
  → windows()
  → Chrome: "GitHub PR #142", "Stack Overflow - auth tokens"
  → Safari: "RFC 6749 - OAuth 2.0"
  → Notes: "Auth refactor plan"

Step 2: Read research notes
  → query(app="notes")
  → "Auth refactor plan: migrate to PKCE flow..."

Step 3: Check what was searched
  → query(app="safari")
  → Last 24h: "PKCE vs implicit flow", "OAuth 2.1 spec"

Step 4: Summarize
  → "Based on open tabs, notes, and search history:
     you're migrating OAuth to PKCE. RFC 6749 is open.
     The auth refactor plan mentions..."
Gemini read across Chrome, Safari, and Notes — three different apps — through MCP. No screenshot of any of them.

​
Free & Freedom

​
Free

Claude CUOpenAI OperatorKernel CLI
Price$20-200/moSubscription$0
VLM cost per action~$0.01~$0.01$0
API dependencyAnthropic APIOpenAI APINone (local)

​
Freedom

Claude CUOpenAI OperatorKernel CLI
Works withClaude onlyGPT onlyAny AI
Who sets safety boundaryAnthropicOpenAIYou
Harness optionsVendor-controlledVendor-controlledRaw / OTP / Enterprise
OfflineNoNoYes
Free: no subscription, no API cost, no vendor dependency. Freedom: any AI, your harness, your rules.

​
Three Harness Levels

The capability is always the same. The control is your choice. 
Level 1: BRAVE MODE (kernel-cli direct)
  No gate. Immediate execution. For development and testing.
  $ kernel-cli ax-press "Delete Everything"  → executes now

Level 2: OTP MODE (niia control)
  Physical harness. OTP from your email. Time-limited.
  $ niia control unlock --scope full --duration 1h
  $ niia control ax-press "Delete Everything"  → after OTP only

Level 3: ENTERPRISE MODE (future)
  Role-based policy. Audit log. Scope-limited per agent.
  Only approved actions. Full compliance trail.
The AI cannot choose its own harness level. You set it. And in OTP mode, no AI — no matter how sophisticated — can bypass it. The OTP lives in your email, outside the AI’s world entirely.

​
Setup

# Install
openclis install niia

# Verify
kernel-cli windows

# Add MCP to any AI
# Claude Desktop: ~/Library/Application Support/Claude/claude_desktop_config.json
# Claude Code: .mcp.json in project root
# Codex: codex.toml [mcp_servers] section
# Gemini: .gemini/settings.json

{
  "mcpServers": {
    "kernel-cli": {
      "command": "kernel-cli",
      "args": ["--mcp"]
    }
  }
}
One install. One config line. Every AI on your machine gets computer use.
Kernel CLI is part of NIIA by Monolex. Free binary. Proprietary source. No subscription. No vendor lock.