MCP is the open standard that lets Claude use real tools. Fine Structure ships a connector with roughly 70 of them, so one chat can build and run an app.
Founders and operators who use Claude and want it connected to a platform that can build, store, publish and keep running.
- A plain-language understanding of what an MCP connector actually is
- A map of what one connected conversation can do: build, data, access, domain, agents, analytics
- The safety model: consent, revocable tokens and rollback versions
MCP, the Model Context Protocol, is the open standard that lets assistants like Claude call external tools instead of writing about them. Fine Structure's connector at https://finestructure.ai/api/mcp attaches from claude.ai settings with one OAuth 2.1 approval and no API keys. Its roughly 70 tools cover app generation, file editing with automatic saved-version rollback, database entities, publishing, custom domains with automatic TLS, access policies and traffic analytics. It also leaves agents that keep working on a schedule, and every grant is revocable.
MCP stands for Model Context Protocol, the open standard for connecting AI assistants to external tools. Why it matters has almost nothing to do with protocols. Ask Claude how to launch a booking site with no connector attached and you get a good plan and a lot of tabs to open. Ask with one attached and Claude starts working inside a real platform.
Connector is the user-facing name for an MCP server your assistant may call. Anthropic runs a connectors directory so people can find them, and the ecosystem is past the experiment stage: Vercel, Neon and Notion all ship connectors. Read that as a signal. When infrastructure companies invest engineering time in a standard, it is settled enough to build on.
For a builder the practical question is narrower. How much of the work between an idea and a working product can happen inside the conversation where the idea appeared? Every handoff costs context, and each retelling loses a detail.
The one line version: MCP is how an AI assistant stops giving advice about your software and starts operating it.
Roughly 70 tools. That sounds heavy until you see the grouping: they follow the life of an application rather than the shape of an API.
What one conversation can do
Intent | What you say | Tools involved | What exists afterwards
Build | Build me a booking app for a small clinic | create_app, file editing | A working app with real screens and a saved version to roll back to
Store data | Each patient needs a name, a phone and a visit date | entities: schemas and records | A schema and its records, defined in the same chat as the screens
Gate access | Only signed-in staff should see the schedule | access policies | Login-gated pages instead of a public URL
Connect domain | Put it on clinic.example, not the platform URL | custom domains, automatic TLS | A live domain over HTTPS, no separate certificate task
Keep it running | Every Monday, summarise last week and message me | create_agent, schedule_agent_task | A scheduled agent reporting to its verified owner on WhatsApp or email
Measure | How much traffic did the booking page get? | get_app_analytics | Traffic figures inside the conversation
Read that as one session, not six errands. A full-stack connector differs from a folder of integrations because the assistant never loses the thread: it knows the schema it wrote when it writes the page that reads from it.
The whole setup is one URL and one approval screen. It takes about a minute.
Adding the connector: In claude.ai open Settings, go to Connectors, choose Add custom connector, and paste https://finestructure.ai/api/mcp. Claude sends you through an OAuth 2.1 sign-in, you approve, and the tools appear in your next message. No config file, no stored API key.
The missing key matters, because API keys are where most integrations quietly go wrong. A key is a long-lived secret that ends up in your clipboard, then a config file you forget, and it keeps working long after you meant it to. OAuth 2.1 replaces it with a grant you can withdraw from your MCP tokens panel in one action.
If you work in Claude Code instead of the web app, there is an open source plugin: finestructure-ai/claude-plugin. It adds /deploy, /fs-status and /fs-domain, covering the shipping loop from your terminal.
This is the question most AI building tools leave unanswered. You get one very good hour, you close the tab, and the software sits there. Nothing checks on it, nothing tells you when the numbers move.
create_agent and schedule_agent_task exist for that gap. An agent is set up inside the conversation that built the app, given a schedule, and keeps working after you leave. It is not a chat window waiting for you. It runs on its own clock.
What makes that useful rather than decorative is the reporting path. An agent can message its verified owner on WhatsApp or email. Verified is the operative word: the destination was confirmed for that account, not typed into a free text field. So scheduled work lands where you read.
Build once, then keep the loop: An app nobody looks at is a project. An app that reports to you on Monday morning is an operation. The difference is one scheduled agent.
Two mechanisms, both deliberately boring. The first is revocation. Everything Claude does runs on a grant listed in your MCP tokens panel, revocable whenever you want. Access ends when you decide, not when a support queue agrees.
The second is rollback. Every change can go back to a saved version, saved automatically as work happens. That changes the arithmetic of letting an AI edit an app real people use. You are not approving a leap of faith, you are approving something reversible.
Now the honest scope, because a connector that oversells itself wastes your afternoon:
What to expect on day one: A free tier with no card, a connector you add in about a minute, roughly 70 tools, a rollback for every change, and a revoke button you never have to hunt for.
MCP, the Model Context Protocol, is the open standard for connecting AI assistants to external tools. A connector is one such tool server, attached with your permission. Claude then acts inside that system instead of describing it.
In claude.ai open Settings, go to Connectors, choose Add custom connector, paste https://finestructure.ai/api/mcp and approve the OAuth 2.1 screen. The tools appear in your next message. Free tier, no card.
Yes. Fine Structure runs one at https://finestructure.ai/api/mcp with roughly 70 tools covering app generation, file editing, database entities, access policies, publishing, custom domains with automatic TLS, traffic analytics and scheduled agents.
No. It uses OAuth 2.1: you approve access once from claude.ai settings and never generate, paste or rotate a key. The grant sits in your MCP tokens panel, revocable at any time.
Anthropic runs a connectors directory and Fine Structure's submissions are in review. Until then, add it as a custom connector with the URL above. It is not being called official, verified or listed, because it is not yet.
Yes. Publishing and custom domains are both connector tools, and TLS is issued automatically, so the certificate is not a separate task. Access policies also put pages behind login.
That is what create_agent and schedule_agent_task are for. An agent set up in the chat keeps running on its schedule after you close the tab, and messages its verified owner on WhatsApp or email.
Every change can roll back to a saved version, saved automatically as work happens. If an edit goes wrong you restore the previous version rather than repairing it by hand.