Key concepts
Authentication
Every connection is authenticated with a project API key — the same key type used to call the external run API. Send it as the x-api-key header on every request. There's no separate MCP-specific credential to manage.
The key determines which organization and project the session operates in. Every tool call is scoped to that organization — there's no way to read or write another organization's data through the MCP server.
Transports
The Anter MCP Server implements the Model Context Protocol over two transports:
| Transport | Endpoint | Use case |
|---|---|---|
| Streamable HTTP (default) | POST/GET/DELETE /v1/external/mcp | The modern MCP transport. What Claude Code, OpenCode, and most current clients use |
| Legacy HTTP+SSE | GET /v1/external/mcp/sse, POST /v1/external/mcp/message | Compatibility path for older MCP clients that haven't adopted Streamable HTTP |
You don't choose a transport explicitly — your MCP client picks the one it supports, and Anter's server implements both against the official MCP SDK.
Sessions
Connecting starts with an initialize call, which returns a session id your client includes on every subsequent request. A session stays open while your client is actively using it and is cleaned up automatically after a period of inactivity — if a session expires, your client reconnects with a fresh initialize call transparently. You don't need to manage this yourself.
Tool categories
Tools are organized by what part of the agent lifecycle they touch:
| Category | What it covers |
|---|---|
| Agent lifecycle | Design an AgentSpec, create or revise a draft, inspect a live agent |
| Running & observability | Trigger a run and read back its result, history, or routing trace |
| Projects & LLM providers | List projects; configure model providers |
| MCP server management | List, add, update, or test the MCP servers connected to your projects |
| Skills & toolsets | Browse your skill catalog and Anter's builtin toolset catalog |
| Utility tools | One-off LLM calls, file analysis, and knowledge base search |
Scoping and permissions
Every tool call carries the organization and project derived from your API key — never a value your client supplies. This means:
- You can only see and modify agents, projects, providers, and MCP servers that belong to your own organization.
- Secrets (LLM provider API keys, MCP server auth headers) are never echoed back in a tool response — you'll see a masked value or no value at all, even after you set one.
- Tools that write data (
add_llm_provider,add_mcp_server,create_draft_agent, and similar) re-verify that any project id you pass actually belongs to your organization before touching anything.
Member-level access
Tool calls authenticate the same way the REST API does — no additional owner/admin gate beyond a valid project API key. Anyone who can call the REST API for your organization has the same access through MCP.