LocalPOV

Development context bridge for AI coding agents. One MCP server, 14 tools.

• Terminal capture — every shell session, auto-captured
• Browser console — errors, warnings, unhandled exceptions
• Network failures — 4xx/5xx responses, CORS errors, slow requests
• Build errors — structured {file, line, col, message} from TS, ESLint, Webpack, Vite, Rust, Go
• Docker logs — container output without leaving your editor
• Screenshots — on-demand browser viewport capture
• Port scanning — what's up, what's down
• System health — memory, CPU, uptime

Your AI agent sees everything happening in your dev environment — without you copy-pasting anything.

Quick start

Claude Code

claude mcp add localpov -- npx -y localpov --mcp

That's it. One command. Restart Claude Code and all 14 tools are available.

Shell integration is installed automatically on first run — restart your terminal to start capturing.

Claude Desktop

Add to claude_desktop_config.json:

• macOS: ~/Library/Application Support/Claude/claude_desktop_config.json
• Windows: %APPDATA%\Claude\claude_desktop_config.json

{
  "mcpServers": {
    "localpov": {
      "command": "npx",
      "args": ["-y", "localpov", "--mcp"]
    }
  }
}

Cursor

Add to .cursor/mcp.json in your project root (or ~/.cursor/mcp.json for global):

{
  "mcpServers": {
    "localpov": {
      "command": "npx",
      "args": ["-y", "localpov", "--mcp"]
    }
  }
}

VS Code / Copilot

Add to .vscode/mcp.json in your project root:

{
  "servers": {
    "localpov": {
      "command": "npx",
      "args": ["-y", "localpov", "--mcp"]
    }
  }
}

Windsurf

Add to your MCP config:

• macOS/Linux: ~/.codeium/windsurf/mcp_config.json
• Windows: %USERPROFILE%\.codeium\windsurf\mcp_config.json

{
  "mcpServers": {
    "localpov": {
      "command": "npx",
      "args": ["-y", "localpov", "--mcp"]
    }
  }
}

Or configure via Windsurf UI: Cascade panel → MCPs icon → Add server.

Any other MCP client

LocalPOV uses stdio transport. Any MCP client that supports stdio works with:

{
  "command": "npx",
  "args": ["-y", "localpov", "--mcp"]
}

Why

AI coding agents are blind. They can read and write files, but they can't see:

• What your terminal just printed
• That the browser console is full of CORS errors
• That fetch('/api/users') is returning 500
• That the Docker container crashed 30 seconds ago
• That TypeScript found 12 errors on build

You end up copy-pasting error messages into chat. LocalPOV fixes this — the agent reads your dev environment directly.

What the AI agent sees

14 tools available via MCP:

| Tool | What it does | |------|-------------| | get_diagnostics | One-call health check: terminal errors, browser console, network failures, port status, memory | | list_sessions | All captured terminal sessions (PID, shell, alive/dead) | | read_terminal | Last N lines from any terminal session, with pagination | | read_command | Output of a specific command by index (supports negative indexing) | | get_errors | Errors across all sessions (JS, TS, Python, Rust, Go patterns) | | search_all | Regex search across all terminal output | | read_browser | Browser console errors + failed/slow network requests | | take_screenshot | Capture browser viewport | | get_build_errors | Structured build errors: {file, line, col, message} | | docker | List containers or read container logs | | tail_log | Read last N lines of any log file | | check_ports | What's listening, what's down | | check_env | Which env vars exist (never exposes values) | | process_health | System memory, CPU, uptime |

How it works

Terminal capture

Shell integration hooks into your shell startup (bash, zsh, fish, powershell). Every terminal session's output is captured to ~/.localpov/sessions/ as plain text logs. The AI agent reads these via read_terminal or get_errors.

This is installed automatically when you run npx -y localpov for the first time.

Browser capture (optional)

Start the proxy to capture browser console and network activity:

npx -y localpov --port 3000

Then open http://localhost:4000 instead of http://localhost:3000 — your app works exactly the same, but the agent can now see console errors, network failures, and take screenshots.

No browser extension needed. Works in any browser.

Dashboard

Open http://localhost:4000/__localpov__/ to see a visual debug panel:

• Apps — detected dev servers with one-click preview
• Terminal — live terminal output stream
• Debug — console errors, network failures, and system health in real-time

Architecture

Terminal sessions          LocalPOV MCP Server          AI Agent
  (bash/zsh/ps)                   |                  (Claude Code,
       |                          |                   Cursor, etc.)
       |  ~/.localpov/sessions/   |   read_terminal         |
       |------------------------->|<------------------------|
                                  |                         |
Browser --> localpov proxy (:4000)|   read_browser          |
              |  injects capture  |<------------------------|
              |  script into HTML |                         |
              |                   |   take_screenshot       |
              |  console + fetch  |<------------------------|
              |  errors captured  |                         |
                                  |   get_diagnostics       |
Docker containers                 |<------------------------|
  |  docker logs                  |                         |
  |------------------------------>|   docker                |
                                  |<------------------------|

Config file

Create .localpovrc in your project root:

{
  "port": 3000,
  "listen": 4000,
  "ports": [3000, 3001, 5173, 8080]
}

Supported shells

| Shell | Capture method | Platforms | |-------|---------------|-----------| | bash | script command + preexec/precmd hooks | Linux, macOS | | zsh | script command + preexec/precmd hooks | Linux, macOS | | fish | script command + fish_preexec/fish_postexec | Linux, macOS | | PowerShell | Start-Transcript + prompt function flush | Windows, Linux, macOS |

Safety

• Output capped at 50KB per MCP response — prevents flooding the AI context
• Environment values never exposed — check_env only reports existence (true/false)
• File access restricted — tail_log only reads from cwd, ~/.localpov/, /var/log/, /tmp/
• Sensitive paths blocked — .ssh, .env, credentials files are never readable
• WebSocket limits — max 50 connections, 64KB message size cap
• Stale data cleanup — sessions >24h and browser data >4h auto-cleaned

Requirements

• Node.js 18+

License

MIT