Helix MCP

Helix supports the Model Context Protocol (MCP) to extend the agent's capabilities. An MCP server exposes a set of tools and prompts that Helix can call as part of its reasoning loop.

What MCP adds

• Access to services that do not have native Helix tools yet.
• Structured prompts and responses from external systems.
• Live, schema-driven tool definitions so the model knows exactly what arguments are required.

Supported server types

• stdio — a local command that speaks MCP over stdin/stdout.
• sse — an HTTP server that streams MCP events.
• websocket — an MCP server over a WebSocket connection.

Configuration

MCP servers are configured in a team-scoped mcp_servers table. Each record contains:

• name — a stable identifier used by the agent.
• transport — stdio, sse, or websocket.
• url or command — how to reach the server.
• headers or env — authentication or runtime variables (stored encrypted).
• enabled — whether the server is active for the team.

Tool manifest

When a conversation starts, Helix builds a manifest of all available tools: built-in Helix tools, team MCP tools, and global integrations. The manifest is cached per teamId so subsequent requests are fast. A HELIX_AGENT_SERVER_URL environment variable can point Helix to a specific agent server instance for development.

Security model

MCP tools run with the same approval gates as built-in tools. A tool that writes, deletes, or calls external APIs is considered risky and triggers the critic unless the conversation autonomy level is set to allow it.

Adding an MCP server

Admins can add servers from the team Settings > Integrations page. Helix validates the server by fetching its tool manifest before enabling it.

Related

• Helix Architecture
• Helix Security