Configure an MCP server

An MCP server (Model Context Protocol) is a tool server an AI agent can call. You set one up, assign it to an agent, and the agent decides which of its tools to call as it reasons. MCP servers are managed on the MCP Servers tab of Settings → Integrations — alongside connectors, the other way to give an agent tools.

Not to be confused with connecting your own AI assistant

This page is about adding a tool server that an agent you build can call. If you want to point your own AI assistant (Claude Code, Claude Desktop, or similar) at Superprocess’s own MCP servers to author or operate artifacts yourself, see Use Superprocess from an AI assistant instead.

Before you start

• You’re an organization owner or admin (MCP servers live on Settings → Integrations).
• You have either a SQLite database to expose, or an MCP server reachable over stdio, http, or sse, with its endpoint and any auth details.

Steps

1. Open the MCP Servers tab. Go to Settings → Integrations and switch to the MCP Servers tab.

   Screenshot placeholderSettings → Integrations — the MCP Servers tabReplace with real capture from app.superagent.studio

2. Add a server. Two ways in:

   • Upload Database — upload a SQLite file and the platform exposes it as a query tool server.
   • Add MCP Server — start from a template (Document, MarkitDown) or go Custom.

3. Set the connection (custom server). Choose the transport — stdio (bundled tools), http, or sse (streaming) — and give the endpoint (for http/sse) plus any auth (for example a bearer token, stored as a secret).

4. Name it and scope it. A short name agents will reference, and a scope — organization-wide, or limited to a project.

5. Check the tools. The platform fetches the server’s tools so you can see what it exposes.

6. Save. The server is registered.

7. Assign it to an agent. Open the agent, go to the Tools & Integrations tab, move the server from Available to Assigned, and click Save Tool Configuration. See Give an agent tools.

Access control is server-side

You don’t grant read-only or read-write per binding. Which tools an agent actually sees, and whether it may perform writes, are enforced server-side by the caller’s role — non-admins get a read-safe subset. So an agent can’t be handed a tool the caller isn’t allowed to use.

Versioning

MCP servers are versioned. Publish a version and pin it on an agent’s assignment, so the tools an agent calls don’t shift underneath a running process.

Troubleshooting

• No tools appear. Check the endpoint and transport; for a custom server, confirm it actually speaks MCP (a trailing-slash mismatch is a common culprit).
• Tools are there but the agent doesn’t call them. Either the agent’s prompt doesn’t say when to use them, or the model isn’t tool-calling-capable. Add an explicit instruction (“look up the PO with lookup_po to verify the vendor”) and confirm the model supports tool calls.
• Tool calls time out. The server is too slow; the per-tool timeout caps at 300 s. Move heavy work off the request path, or simplify the tool.

Related

Give an agent toolsAssign connectors and MCP servers to an agent.

Build an integrationConnectors — the other kind of agent tool.

Create an agentThe agent these tools attach to.

AI agentsHow agents use tools.