Every AI coding tool speaks MCP now. Claude Code, Cursor, Windsurf, VS Code Copilot, Codex CLI, Gemini CLI, OpenClaw — they all support the Model Context Protocol.

But configuring MCP for each tool means editing JSON files in different locations, managing tokens, and remembering which tool uses which config format. We built a single command that handles all of it.

The command

npx @inariwatch/mcp init

That's the entire setup. No install, no global CLI. npx runs it directly.

What happens when you run it

Step 1 — Detect

The init script scans your machine for every supported AI tool:

Claude Code — checks if claude is in PATH
Cursor — checks for ~/.cursor/ config directory
Windsurf — checks for ~/.windsurf/ config directory
VS Code + Copilot — checks if code is in PATH
Codex CLI — checks if codex is in PATH
Gemini CLI — checks if gemini is in PATH
OpenClaw — checks if openclaw is in PATH or ~/.openclaw/ exists

Each tool gets a green check or a skip. No manual selection needed.

Step 2 — Authenticate

A browser window opens with a device code. You approve it on InariWatch, and the CLI receives your API token. No copy-pasting tokens from dashboards.

Authenticating via browser...
Code: ABC123
Opening: https://app.inariwatch.com/cli/verify?code=ABC123

If you already have a token, skip the browser flow:

npx @inariwatch/mcp init --token YOUR_TOKEN

Step 3 — Configure MCP

Here's where each tool gets different treatment. MCP config locations vary across tools:

Claude Code — configured via claude mcp add CLI command
Cursor — writes to ~/.cursor/mcp.json
Windsurf — writes to ~/.windsurf/mcp.json
VS Code — writes to .vscode/mcp.json
Codex CLI — configured via codex mcp add CLI command
Gemini CLI — configured via gemini mcp add CLI command
OpenClaw — configured via openclaw mcp set or writes to ~/.openclaw/openclaw.json

The init script handles all of this — correct JSON structure, correct auth headers, correct transport protocol. One token, seven configurations.

Step 4 — Install the Capture SDK

If you're inside a Node.js project, init also installs @inariwatch/capture and wires it into your framework:

Next.js — wraps next.config with withInariWatch, creates instrumentation.ts
Nuxt 3 — adds the module to nuxt.config
Remix / SvelteKit / Astro / Vite — adds inariwatchVite() to plugins
Express / Fastify / Node — installs the SDK with instructions

Skip this step with --no-capture if you only want MCP.

Step 5 — Link GitHub

If the gh CLI is installed and authenticated, init offers to link your GitHub account. This enables code intelligence (semantic search across your repos), PR risk assessment, and AI remediation that reads your actual source code.

What your AI tool gets

Once configured, every detected AI tool has access to 25 MCP tools:

Query — query_alerts, get_status, get_uptime, get_build_logs, get_error_trends

Diagnose — get_root_cause, assess_risk, search_community_fixes, get_postmortem, ask_inari

Fix — trigger_fix, rollback_deploy, silence_alert, acknowledge_alert, reopen_alert

Monitor — create_uptime_monitor, run_health_check, reproduce_bug, simulate_fix, verify_remediation, search_codebase

Plus 4 live resources (critical alerts, recent alerts, status overview, active remediations) and 7 prompt workflows.

The MCP server runs at mcp.inariwatch.com using Streamable HTTP. Auth is Bearer token with SHA-256 hashing. Rate limits apply per tool tier — 200/min for queries, 30/min for analysis, 5/min for execution.

The design principle: sampling-first

Four of the analysis tools — get_root_cause, assess_risk, ask_inari, and simulate_fix — don't call any AI on the server. Instead, they gather context (alerts, code, recordings, community fixes) and return it to your AI tool's LLM for analysis.

Your AI editor already has a model running. Why pay for a second inference call on the server? InariWatch provides the data, your tool provides the intelligence.

Only trigger_fix uses server-side AI — because remediation requires a multi-step pipeline (diagnose, read code, generate fix, self-review, push, CI, PR) that runs asynchronously.

Why one command matters

Before init, connecting InariWatch to three tools meant:

1. Go to Settings, copy your token 2. Open Claude config, paste the JSON 3. Open Cursor config, paste different JSON 4. Open VS Code config, paste yet another format 5. Remember to update all three when rotating tokens

Now it's one command. When we add support for a new AI tool, your existing setup doesn't change — just run init again and the new tool gets configured alongside the rest.

Try it

npx @inariwatch/mcp init

Works on macOS, Linux, and Windows. Detects your package manager automatically (npm, yarn, pnpm, bun).

npm · Docs · GitHub

7 AI tools, one command: how npx @inariwatch/mcp init works