Skip to main content
Helix speaks the Model Context Protocol (MCP). It ships no MCP server of its own: the servers you configure are the ones it uses, and their tools reach the model directly, next to Helix’s own tools.

Add a server

List your servers in an mcp.json file, in the mcpServers shape other MCP clients use:
mcp.json
A local server runs from command and args; a remote one is reached at url. The next Helix session loads them.

How the model sees the tools

Each tool is offered to the model as mcp__<server>__<tool>, for example mcp__docs__search. The model calls it like any other tool.
Leave a server’s exposure field out, or set it to "direct". A server set to "codemode" or "deferred" has no tools Helix can call, and Helix warns about it.

Manage servers with /mcp

Type /mcp in a Helix session to open the MCP manager. It lists every server you configured with its state (for example connected · 4 tools, needs sign-in, failed or disabled), its exposure, and where it is configured (for example project). Servers that need you come first. With no servers configured, it tells you which files to add them to. Pick a server to act on it: Sign-in applies to remote servers that have no Authorization header and ask for one when Helix connects: Helix signs you in with OAuth and stores the credentials for you. You can also run an action directly: Leave out the server name and Helix picks the only one that fits, or asks you which one. When a server fails to connect or needs a sign-in at the start of a session, Helix tells you once and points you to /mcp.

Where MCP tools are available

Project context

The other thing every session loads from your project: AGENTS.md / CLAUDE.md.