MCP servers —
one URL. every agent.

MCP servers implement the Model Context Protocol — an open standard that lets an AI agent call external tools and APIs mid-conversation without you writing custom integration code for every model. When Claude, Cursor, or any MCP-compatible agent connects to an MCP server, it can read your database, create a GitHub issue, send a Stripe invoice, or query Notion, all from inside a single chat. On Aerostack, hosted mcp servers run on Cloudflare's edge, so there are no servers to manage, no cold-start delays, and each tool call executes close to the requester. Think of an MCP server as a capability layer that any compatible AI agent can call — you build or install it once, and every agent that connects to your workspace can use it immediately.

These are three distinct primitives that a workspace composes together. An MCP server is a hosted tool server that implements the Model Context Protocol — it exposes a set of callable tools (like search_repos or create_invoice) to any AI agent. A Skill is an agentic workflow template: a pre-built multi-step sequence (code review, summarize document, triage issue) that the agent can invoke as a single named action. An Edge Function is a TypeScript function deployed at the Cloudflare edge — it runs your custom logic, reads from your databases or KV stores, and can be wrapped as a tool that any server or skill can call. You install all three into the same workspace and they are all auto-namespaced behind your single workspace URL; the AI model cannot distinguish which layer is handling its request.

An Aerostack workspace is a single authenticated URL that acts as a unified gateway to every MCP server, skill, and function you have installed. Instead of pasting a separate endpoint for Stripe, another for GitHub, and another for your own custom server into your Claude Desktop or Cursor MCP config, you paste one workspace URL and get all of them at once. The platform handles authentication, secret injection, and routing behind that URL, so your agent sees a merged tool catalog from every server in the workspace. When you add or remove a server, the change is immediately reflected — no config update required on the client side. Secrets like API keys are encrypted at rest and never exposed to the model.

When your workspace composes multiple MCP servers, tool names can collide — two servers might both define a tool called search or create. Aerostack prevents this by automatically prefixing every tool name with its server slug, separated by a double underscore. A tool called search_repos on the github-mcp server becomes github-mcp__search_repos; a tool called query on the postgres-mcp server becomes postgres-mcp__query. The AI model sees the prefixed names in its tool list and uses them exactly — no ambiguity, no conflicts, regardless of how many servers you compose. This namespacing happens automatically; you do not configure it.

The setup is a single paste. Open your Aerostack workspace, install the MCP servers you want, then copy the generated workspace URL into your AI client's MCP configuration — Claude Desktop, Cursor, ChatGPT, Claude Code, Gemini, CrewAI, or any other MCP-compatible tool. From that point on, every tool from every installed server appears in the agent's tool list automatically. There is no per-server configuration to repeat. If you later add another server to the workspace, it shows up in all connected clients without any further changes. The workspace URL stays the same regardless of how many servers you add or swap out. For OpenAI and Gemini formats, the workspace also auto-generates a /openai-tools and /gemini-tools endpoint from the same URL.

The Aerostack mcp marketplace has 254 hosted MCP servers across 15 categories: developer tools, productivity, payments, databases, communication platforms, CRM, analytics, automation, security, web search, version control, e-commerce, cloud storage, AI and LLMs, and more. Community creators publish new servers regularly, and each listing shows what tools the server exposes, what credentials it needs, and whether it is free or paid. You can browse and search the full catalog on the Explore page — filter by category, popularity, or compatibility. Installing a server from the mcp registry adds it to your workspace in one click; your secrets are stored encrypted and scoped to your workspace so they are never shared with other users or visible to the model.

Yes — private mcp server hosting is one of the core use cases. You write your MCP server logic, deploy it to Aerostack, and it runs on Cloudflare Workers at the edge: near-zero cold start, global distribution, and no infrastructure to maintain. Private servers are visible only to your workspace and are never listed in the public mcp marketplace. This is how teams expose internal APIs, proprietary databases, or custom business logic to their AI agents without publishing anything publicly. If you later want to share a server with the community, you can flip it to public and it becomes discoverable in the mcp registry. Full details on private hosting are on the Private MCP page.

Aerostack has a free tier that includes 500,000 AI tokens per month — enough for meaningful development and light production use. Usage beyond the free tier is metered against your plan. If you bring your own model API key (BYOK), the platform markup drops to zero, so you only pay the underlying provider's rate. Marketplace servers that carry a usage fee (set by the creator) show the price clearly on the listing. There are no per-server seat fees — you pay for execution, not for the number of servers you install.

Build your MCP server on Aerostack — either from scratch using the server framework or by proxying an existing API — then set its visibility to public from your workspace settings. Once published, it appears in the mcp marketplace catalog and any Aerostack user can install it into their own workspace. Publishing is free; to charge for usage, set a price on your agent endpoint (pay-per-call). You keep full control over the server code and can update it at any time; all installed workspaces get the update automatically. Browse existing published servers on the Explore page to get a sense of what categories are popular and what gap your server might fill.

Yes. All credentials — API keys, OAuth tokens, connection strings — are encrypted at rest using AES-256 before being stored. At runtime, secrets are injected into the server process on the edge and are never returned to the AI model or logged in plaintext. The model only ever sees tool names, descriptions, and the structured outputs of tool calls. Each workspace has its own secret scope — secrets are not shared across users, workspaces, or servers. If you revoke a secret from your workspace settings, it is invalidated immediately across every server that used it.

If you only need a single remote mcp server and your AI client supports direct server connections, running it standalone is simpler — one less layer in the stack. Aerostack's workspace model earns its place when you need more than one server, because the workspace composes them behind a single URL instead of forcing you to manage N separate endpoints in every client config. The workspace also adds encrypted secret storage, access controls, and the ability to mix mcp marketplace servers with your own private ones in one place. If you are a solo developer doing quick experiments with a single community server for Claude or Cursor, start standalone and migrate to a workspace when you hit the multi-server wall.