MCP servers
What is MCP #
The Model Context Protocol (MCP) is an open standard that lets AI agents discover and call external tools. An MCP server exposes a set of tools that agents (ChatGPT, Claude, Cursor, etc.) can invoke to interact with APIs, databases, or any external service.
What Bump.sh offers #
Bump.sh lets you create MCP servers by declaring workflows in a simple document. You describe the API calls and their sequencing, and Bump.sh generates a fully hosted MCP server that executes them.
This approach is declarative: you define what the server should do, not how to run it. There is no code to write, no server to deploy, no infrastructure to manage. Because every execution follows your workflow definition exactly, the behavior is deterministic and predictable, which is critical when AI agents interact with real APIs.
The data plane #
All API calls are executed by an isolated component called the data plane. The AI agent never calls your APIs directly. Instead, it invokes a tool on your MCP server, and the data plane runs the corresponding workflow: resolving secrets, executing HTTP requests, and returning only the declared outputs to the agent.
This architecture keeps credentials and sensitive API responses away from the LLM. See Security for details.
Workflow specifications #
Workflows can be written in two formats:
- Flower: a lightweight specification designed by Bump.sh. Ideal for small projects, quick prototyping, or workflows calling APIs for which you don’t have an OpenAPI document.
- Arazzo: the workflow specification from the OpenAPI Initiative. Better suited for complex projects that need to reference existing OpenAPI documents.
Both formats support multi-step sequences, conditional logic (retry, goto, end), runtime expressions, and secrets.
Get started #
- Create an MCP server from your dashboard.
- Deploy a workflow document that describes the API calls your server can perform.
- Share the server URL so that end-users can add it to their AI tool.
Go further #
- Secrets: store API keys and tokens so they are never exposed in your workflow documents.
- Access management: control who can use your MCP server.
- Debug sessions: inspect execution traces to troubleshoot workflows step by step.
- Runtime expressions: reference dynamic data between steps.
- Security: understand the architecture and data handling guarantees.
MCP servers are in closed beta. Contact us to be among our first users.