# MCP Server

Source: https://www.jawsdeploy.net/features/mcp

Connect Claude Code, Claude Desktop, Cursor or Codex to Jaws, using the AI subscription you already have. No extra fees. No model lock-in. No new vendor to evaluate.

## Your AI assistant, your deployments, your bill.

Jaws ships with a built-in Model Context Protocol server. Point Claude Code, Claude Desktop, Cursor, Codex Desktop, Codex CLI or another MCP client that supports remote servers at your Jaws workspace, and the AI can create projects, assemble deployment pipelines from step templates, read deployment logs, author new templates and set variables on your behalf.

You pay your AI provider once. We don't charge per token, we don't proxy your model calls, and we don't tie you to a specific vendor. If you switch from Claude Code to Cursor next month, the same MCP account works there too.

## What you can do today

- **Build deployment pipelines from a description**: Tell your AI what you need to deploy and how. It creates the project, adds the right steps from your template library, and fills in each step's properties. Everything lands in Jaws ready to review.
- **Diagnose failed deployments**: Ask about a failed deployment. The AI finds it, reads its log (optionally only errors and warnings) and the settings of the steps involved, and explains in plain English what broke and how to fix it.
- **Manage variables in bulk**: Paste a config dump or environment definition, ask the AI to set the corresponding Jaws variables, and watch it happen. Secret variables are deliberately blocked: you still set those by hand.
- **Navigate and organise by conversation**: List my workspaces. Move all staging projects into a folder called Staging. Lookups and reorganisation that used to take six clicks become one sentence.

## A three-step setup

1. **Create an MCP account.** In Jaws, open **Settings → AI assistants → Create MCP account**, pick a workspace and your AI assistant. Jaws creates a service account limited to that workspace and generates its API key.
2. **Add Jaws to your AI client.** The same page shows the setup for your assistant with the credentials filled in. The [quickstart guide](https://www.jawsdeploy.net/guides/mcp-quickstart) has every client's setup too.
3. **Start asking questions.** Your AI now sees the workspace's projects, step templates, variables and deployment logs.

## Why this approach

**We could've charged you per token.**

Most deploy tools that ship AI features quietly become token resellers. They wrap a model provider, mark up the inference, and bake it into a new pricing tier. That's fine for them and bad for you — your costs scale with conversations, you're locked into whichever model they chose, and your spend on AI is now split across two bills.

**We chose MCP instead.**

By exposing Jaws as an MCP server, we let your existing AI subscription do the work. You bring your Claude, Cursor, or Codex plan. We bring the deployment surface. No inference markup. No new vendor to evaluate. When models get cheaper and better — which they will — the savings go straight to you.

## Built so the AI can't break things on its own

- **Deploying takes a preview first.** Triggering a deployment is split into `preview_deploy` (a dry run) and `trigger_deploy`, which only accepts the short-lived token the preview returns. The assistant is told to wait for your go-ahead before it deploys.
- **No secrets in transcripts.** Secret variable values are masked when read and rejected when written. Secrets are set by hand in the Jaws UI and never travel through an AI conversation.
- **Every call is permission-checked.** Tool calls go through the same workspace permission checks as the REST API. A service account that can't manage a workspace can't touch it via MCP either.
- **Revocable in a minute.** Delete the MCP account and the assistant loses access within a minute.

- [Read the quickstart](https://www.jawsdeploy.net/guides/mcp-quickstart)
- [Start free trial](https://app.jawsdeploy.net/signup)