Every agent starts with a base set of tools: run commands, read and write files, search directories. That covers a lot. But some tasks need more. Web search, database queries, GitHub access, custom APIs.
That's what MCP servers are for.
What MCP servers do
MCP (Model Context Protocol) servers are tool providers. Each server exposes a set of capabilities that agents can call during a run. An omnisearch server gives agents web search and AI-powered research. A GitHub server gives agents access to repos, issues, and pull requests. A SQLite server gives agents database access.
You configure MCP servers once, then attach them to any agent that needs those tools.
Built-in templates
Go to MCP Servers and click Add Server. Two templates are ready to go:
mcp-omnisearch. Web search, GitHub search, AI-powered research, URL extraction, and content summarization. Hooks into third-party search services (Tavily, Brave, Kagi, Exa, Perplexity, GitHub) via your own API keys. Add the keys you have, skip the ones you don't.
mcp-sqlite-tools. SQLite database management. Create, query, and modify databases inside the sandbox.
Pick a template, add your API keys if needed, and the server is ready to attach to agents.
Custom servers
Click Custom to configure any MCP-compatible server. You need:
- Name and optional description
- Type: stdio (runs as a subprocess) or HTTP (connects to a URL)
- For stdio: the command and arguments to launch the server
- For HTTP: the server URL and optional headers
Any server that speaks the MCP protocol works. If you've built a custom tool server for your own APIs, plug it in here.
Secrets management
MCP servers often need API keys. When you add a server from a template, Svortie shows which secrets are required and whether they're already configured. Missing a key? Add it inline, right from the setup flow.
Secrets are stored securely and injected into the sandbox at runtime. The agent never sees the raw key in its prompt.
Attaching to agents
In your agent's settings, under MCP Servers, pick which servers this agent should have access to. The agent can then call any tool those servers provide.
If you use AI generation to create agents, the generator looks at your available MCP servers and picks the relevant ones automatically based on the task description.
What agents see
During a run, MCP server tools appear as callable functions alongside the built-in tools. The agent decides when to use them based on the task. You'll see every MCP tool call in the run logs, so there's full visibility into what the agent did and why.
Try it
Add the omnisearch MCP server at MCP Servers, then create a research agent that references web search in its task. Watch the run logs to see it pull real data from the internet.