MCP Tools

Sidebar → Tools → New

Model Context Protocol (MCP) lets agents call external tools through a standard protocol. sbot supports stdio, HTTP, and SSE transports, with built-in presets, global servers shared across agents, and per-agent servers.

Adding a Server

• stdio — command + args (e.g. npx -y some-mcp-package); environment variables can be configured per server
• http — remote MCP endpoint URL + optional headers
• sse — remote URL + optional headers (for hosted MCP services)

Configuration

• Built-in presets — Playwright, Markitdown, Exa, and local built-in tool groups are listed from the same MCP screen
• Global servers — shared across every agent
• Per-agent servers — open an agent → MCP tab to enable global providers or add agent-only MCP servers
• Auto-restart — failed stdio servers are automatically respawned
• Lazy start — servers boot only when an agent that uses them runs
• Tool timeout — override the per-call timeout for slow tools
• Prompt / Resource tools — optionally expose MCP prompts and resources through generated helper tools

Usage

Once attached to an agent, MCP tools are advertised to the model on every turn. The Web UI can inspect a provider's tools, prompts, resources, and resource templates, and tool results are fed back into the conversation.

Tips

• For local commands that need a Node toolchain, use the npx -y form to avoid pre-install hassle
• Use HTTP or SSE transport when the MCP server is remote or shared across multiple sbot instances
• Sensitive secrets in env vars are masked in logs