On this product
MCP servers
The Model Context Protocol (MCP) is an open standard for exposing tool catalogs to AI agents. When you connect an MCP server to Anter, it discovers every tool the server exposes automatically — you don't configure each one by hand.
Looking for Anter's own MCP server?
This page covers connecting a third-party MCP server to Anter, so your agents can call its tools. If you want to connect an MCP client (Claude Code, Claude Desktop, OpenCode) to Anter and manage your agents conversationally, see the Anter MCP Server section instead.
When to use an MCP server vs an HTTP tool
| HTTP tool | MCP server | |
|---|---|---|
| Best for | A single webhook or API endpoint | A collection of related tools from one system |
| Setup | Configure each tool individually | Connect once — all tools are discovered automatically |
| Maintenance | Update each tool schema manually | Tool catalog updates when the server updates |
| Authentication | Secret header per tool | Managed at the server connection level |
Use an HTTP tool when you need one specific endpoint. Use an MCP server when a system (your internal platform, a database, a SaaS product) exposes many related capabilities and you want the full catalog available.
Connecting an MCP server
MCP servers are connected from Agent Builder → MCP → Catalog (or a custom server on that page). A connected server's tools are available to flows in the projects you attach. The Integrations page does not add MCP servers; it can show a server that is already connected.
Each server connection requires:
Name and description — how the server is identified in Anter.
Transport — how Anter connects to the server:
| Transport | Use case |
|---|---|
streamable_http | A server that accepts HTTP connections and streams responses. The most common choice for hosted servers |
sse | A server that uses Server-Sent Events. Use for servers that don't support streamable HTTP |
stdio | A local process Anter spawns directly. For development and self-hosted tools |
URL — the endpoint of the server (for streamable_http and sse transports).
Authentication — MCP servers that require OAuth or token authentication store credentials encrypted at the connection level. The credentials are injected into requests automatically.
How tool discovery works
When Anter connects to an MCP server, it queries the server for its tool catalog — the list of tools it exposes, with names, descriptions, and input schemas. This catalog is cached and made available to sub-agents.
If the server's tool catalog changes, Anter picks up the updates on the next connection. The tools available to sub-agents reflect the current state of the server.
Attaching MCP toolsets to a sub-agent
Connecting an MCP server makes its tools available to the project — it doesn't automatically give every sub-agent access. You control which sub-agents can call which server's tools.
Open the sub-agent editor by selecting a node on the canvas. The MCP toolsets selector lists the servers connected to this project. Select the servers whose tools this sub-agent should have access to.
Like HTTP tools, keep the tool surface narrow per sub-agent. A sub-agent that can call every tool in a large MCP server catalog has more surface area for unexpected behavior than one given access to a specific, relevant subset.
MCP tools and permission mode
MCP tools are subject to the same permission mode and risk category rules as HTTP tools. If the server exposes tools with destructive or external side effects, make sure the sub-agent's permission mode is set appropriately.
How to vet MCP servers
When evaluating a third-party or internal MCP server for your project, verify these five criteria before attaching it to production sub-agents:
- Narrow scoping: A well-designed server exposes small, single-purpose tools rather than an unbounded catalog. Sub-agents given access to dozens of unrelated tools experience higher prompt overhead and greater routing ambiguity.
- Deterministic parameter schemas: Verify that tool parameters specify strict data types, required fields, and bounded enums to prevent prompt injection and malformed tool executions.
- Authentication & secret isolation: Ensure the server authenticates via header tokens or OAuth stored encrypted at the connection level, and never reflects credentials in error logs or execution payloads.
- Least-privilege permission modes: Classify each tool's risk profile. Tools that perform irreversible writes, outbound communications, or data deletion should run under audited permission modes with explicit step budgets.
- Clear semantic descriptions: Tools must have concise, unambiguous descriptions so the LLM knows precisely when, why, and how to invoke them without guess-and-check loops.
Server types
When creating an MCP server connection, Anter asks for a server type. This is a hint that helps Anter configure the connection correctly:
| Type | What it's for |
|---|---|
vector_retrieval | Knowledge base or semantic search servers |
postgresql | Direct database query tools |
web_tools | Web search, fetch, and browsing tools |
custom | Anything else |
The server type does not restrict what tools are discovered — it's informational and helps with connection defaults.