What your AI agent can write, create, and deploy right now

Stripe has 25 MCP tools and your agent can process a refund through them. WordPress shipped an MCP adapter and it can't edit a post. Klaviyo's agent integration can do a lot, but it still can't send a campaign. The gap between the tools that let agents act and the ones that don't is getting wide enough to change how you pick your stack.

We went through 15 MCP servers and checked what agents can actually write, create, modify, and send today. Not what's announced, not what's in beta. What works, and where the chain breaks.

Payments

Stripe. 25 tools. Create customers, set up subscriptions, issue invoices, process refunds. The full payment lifecycle, no human in the loop. They also launched a Machine Payments Protocol in March 2026 for agent-to-agent transactions, which is its own rabbit hole.

E-commerce

Shopify has five MCP servers covering storefront, catalogue, checkout, customer accounts, and dev tooling. The community shopify-mcp package adds 31 tools on top of that: full CRUD on products, orders, inventory, and discounts through the Admin GraphQL API. So your agent spots a conversion drop in analytics and checks whether a product went out of stock, same session, no tab switching.

Email

Resend's MCP composes and sends emails from Claude Code, Cursor, or Codex. Not drafts them for your review. Sends them. Mailgun does the same, individual or batch, transactional and marketing.

Websites

Webflow shipped an official MCP in February 2026. Create responsive layouts, update CMS content in bulk, modify styles, fix broken links, manage SEO metadata. The catch: you need their Bridge App open for design changes, and it can't create new pages yet. But for CMS work and content updates, it's solid.

Databases

Supabase handles tables, SQL queries, edge functions, auth, storage, and TypeScript type generation, all through natural language. Over 832,000 downloads at this point. Neon has 20+ tools for serverless Postgres with database branching and schema management.

The rest of the stack

GitHub does issues, PRs, releases, and code review. Vercel handles deployment and environment variables. Linear and Notion both have official read/write MCP servers. Slack searches, summarises, and sends messages. Gmail does search, drafts, and sending with scoped OAuth.

And for anything without a native MCP, Zapier connects to 8,000+ apps with 40,000+ actions.

Some are still lagging behind

Not everything is there yet, and some of the larger players are still deciding how to approach things.

WordPress. The official MCP adapter landed in WordPress Core 6.9 in February 2026, which sounds great until you realise it ships with exactly three abilities, all read-only: get site info, get user info, get environment info. Write operations are technically possible through the Abilities API, but plugin developers have to register them explicitly, and the official guidance is "prefer read-only abilities for public MCP endpoints." For a platform running 40% of the web, that's a cautious and probably correct call. But it means your agent can read your WordPress site and can't touch it.

Klaviyo. Can create campaigns, templates, profiles, subscribe and unsubscribe contacts. The expanded Anthropic integration dropped literally this week. The agent handles the setup and the final send stays in the Klaviyo UI, leaving the verification step with you. Reasonable for email, where a bad send is hard to undo.

HubSpot. API rate limits make agent usage painful at any real volume. Fine for occasional queries, not for an agent that's checking things every few minutes.

If your workflow depends on one of these for a critical write step, either use Zapier as a bridge or plan around it.

What a closed loop looks like

All off-the-shelf MCP servers, no custom code:

  1. Analytics spots a bounce rate spike on the pricing page
  2. Supabase confirms the page data is rendering correctly
  3. GitHub creates a PR with the fix
  4. Vercel deploys the update
  5. Slack notifies the team
  6. Linear closes the ticket
  7. Analytics confirms the bounce rate dropped the next day

Every step is the agent acting through an MCP server. Six months ago, every step was a tab and a manual click.

What to check before you buy

When you're evaluating a tool for your stack, you're now choosing it for two users: you and your agent. Worth checking:

MCP server. Does it have one? Check PulseMCP (14,000+ servers) or the Glama registry (23,000+). If yes, your agent can operate it. If no, you're the integration layer, and that tax compounds.

Write access. Read-only MCP is useful for monitoring but can't close a workflow. Check whether the tools include create, update, and delete operations.

Rate limits. If the API throttles at 100 requests per hour, your agent will burn through that in minutes. Look for X-RateLimit headers and documented burst allowances.

Auth model. API keys and OAuth work cleanly with agents. Session-based or cookie auth doesn't.

Webhooks. Your agent needs to know when things change without polling every 30 seconds. HMAC-signed webhooks are the gold standard.

The thing is, tools that have thought about agent access tend to have thought about the other things too: clean documentation, predictable error responses, good rate limit transparency. The correlation is surprisingly strong.

The numbers

9,400+ MCP servers as of April 2026, up from 1,200 in early 2025. 78% of enterprise AI teams have at least one MCP-backed agent in production. 97 million monthly SDK downloads.

The gap between what your agent can read and what it can act on is closing fast. Check your stack.