OAuth 2.1 and remote transport
Streamable HTTP servers with OAuth 2.1, so users connect from Claude, ChatGPT, Cursor or VS Code and approve access on your own sign-in page. Bearer tokens for headless agents.
MCP Server Development
Let your users ask Claude, ChatGPT or Cursor to work with your product. We design and build Model Context Protocol servers with OAuth sign-in, scoped permissions and tools that are safe to hand to an AI, and we run three of them in production in our own products.
What we build into every MCP server
Tools an AI can use well
Tool names, descriptions and input schemas designed for models, not just mirrored from your REST API, so assistants pick the right tool and fill it correctly.
Scoped, revocable access
Each connection only sees the areas the user allows, and they can disconnect at any time. Access is logged so your team can see what an assistant changed.
Safe destructive actions
Deletes and other irreversible tools never act on the first call. They return a warning and a confirmation token, so the user always has the final say.
Structured results
Responses return the fields an agent needs, such as extracted codes, links or IDs, instead of raw HTML or long text the model has to parse.
Built on your existing API
We add the MCP layer next to the backend you already have, in Kotlin/Ktor or Node.js/TypeScript, sharing its auth, permissions and data.
MCP servers we run in production
These are our own products, built and operated by BlanketByte. Each one works inside AI assistants today.
Piccola
A calm personal organizer. Users connect Claude, ChatGPT, Cursor or VS Code and choose which areas, like routines, budget or notes, each assistant can see. piccola.app
Gancio
A CRM for consultants and sales reps. Ask your assistant who to call back, log a call or update a contact, with revocable access and API tokens. gancio.app
Poofbox
Disposable inboxes for AI agents. An OAuth 2.1 MCP server that creates an address, waits for the email and returns the code or link. poofboxmail.com
How a project works
- Map the use casesWe list what your users would ask an assistant to do in your product, and which data it must never touch.
- Design the toolsWe turn those use cases into a small set of well-described tools and agree on permissions and confirmations.
- Build and connectWe build the server with OAuth, test it in Claude, ChatGPT and Cursor, and deploy it next to your API.
- Run and improveWe watch how assistants use the tools, fix the ones they misuse and add new ones as your product grows.
Frequently Asked Questions
What is an MCP server?
MCP (Model Context Protocol) is an open standard for connecting AI assistants to other software. An MCP server exposes your product as a set of tools and data that Claude, ChatGPT, Cursor and other MCP clients can use on a user's behalf, after the user signs in and grants access.
Which AI assistants can connect to an MCP server?
Claude (web, desktop and Claude Code), ChatGPT through custom connectors, Cursor, VS Code and any other client that supports remote MCP servers. Our own servers are used from all of these.
Do I need to rebuild my backend?
No. The MCP server sits next to your existing API and reuses its authentication, permissions and business logic. We build it in Kotlin/Ktor or Node.js/TypeScript.
How do you keep it secure?
Users sign in with OAuth 2.1 on your own login page, each connection is limited to the areas they approve, access can be revoked at any time, and irreversible actions require an explicit confirmation step.
Can you also build the AI features inside my app?
Yes. For Piccola we built both: an in-app assistant that proposes every change as a card the user confirms, and an MCP server for people who prefer their own assistant. See AI & MCP integrations.
Put your product inside the assistants your users already use
Tell us what your users should be able to ask for. We will map the tools, scope the work and suggest where to start.