MCP servers
Every MCP server, local or remote, works with your key in every app that speaks MCP.
Every MCP server works with your Evolved key, in every app that speaks MCP. The servers give the model its tools (your files, GitHub, a database, Figma, Blender); the app runs them, and the model behind it decides when to call them. Point the app at evolved.to and that model is ours: tool calls, their results and streaming go through exactly as they would directly.
How it works
Your app starts its MCP servers
Local servers run on your machine; remote ones are reached at their URL. Nothing about them changes for evolved.to.
It sends your prompt and the tools to evolved.to
The tool list travels with the request, on Messages, Chat Completions or Responses, whichever your app uses.
The model calls a tool, your app runs it
The answer asks for a tool with its arguments. Your app calls the MCP server, sends the result back, and the model carries on, as many rounds as the task needs.
Every kind of MCP server
Because the app handles the connection, every MCP transport works the same on your key:
| Transport | What it is | Works |
|---|---|---|
| stdio | A local program your app starts, such as uvx blender-mcp or npx servers | ✓ |
| Streamable HTTP | A remote server at a URL, the current standard for hosted servers | ✓ |
| SSE | Older remote servers that stream over server-sent events | ✓ |
| OAuth sign-in | Remote servers that ask you to sign in (GitHub, Notion, Linear…); your app does the sign-in | ✓ |
Tools, resources and prompts from a server all reach the model, and so do images a tool returns, on models that read images.
Apps that speak MCP
Connect the app to evolved.to with its guide, then add servers where the app keeps them:
| App | Where MCP servers go | Connect it |
|---|---|---|
| Claude Code | .mcp.json in your project, or ~/.claude.json | Guide |
| Claude Code in VS Code & Cursor | The same files as Claude Code | Guide |
| Cursor | .cursor/mcp.json, or Settings → MCP | Guide |
| VS Code | .vscode/mcp.json (agent mode) | Guide |
| Codex | ~/.codex/config.toml, under [mcp_servers] | Guide |
| Cline | MCP Servers → Configure | Guide |
| Roo Code | MCP settings, mcp_settings.json | Guide |
| n8n | The MCP Client Tool node, inside an AI Agent | Guide |
| Factory Droid, OpenClaw, Hermes Agent, Open Cowork | Each app's MCP settings | Guides |
| Blender | Through any app above, with the Blender MCP add-on | Guide |
Add a server
The same two servers in each app: Blender, a local stdio server, and GitHub, a remote HTTP one. Restart the app (or reload its MCP list) after saving.
{
"mcpServers": {
"blender": { "command": "uvx", "args": ["blender-mcp"] },
"github": { "type": "http", "url": "https://api.githubcopilot.com/mcp/" }
}
}{
"mcpServers": {
"blender": { "command": "uvx", "args": ["blender-mcp"] },
"github": { "url": "https://api.githubcopilot.com/mcp/" }
}
}{
"servers": {
"blender": { "command": "uvx", "args": ["blender-mcp"] },
"github": { "type": "http", "url": "https://api.githubcopilot.com/mcp/" }
}
}[mcp_servers.blender]
command = "uvx"
args = ["blender-mcp"]Good to know
Pick a model that is good with tools: every Claude model and GPT-5 and up handle long tool chains well. Tool calls count like any other tokens on your Usage page; a server that returns a lot of text makes requests larger.