MCP servers and subagents

MCP servers

Where they go. .mcp.json at the repository’s root (~/.agents/mcp.json at user scope), in the common mcpServers format:

.mcp.json

{ "mcpServers": { "docs": { "command": "npx", "args": ["-y", "docs-mcp"] } } }

Agent	Gets
Claude Code, Copilot, Devin	nothing written: they read .mcp.json where it is
Codex	[mcp_servers.<name>] in .codex/config.toml
Gemini CLI, Qwen Code	mcpServers in their settings.json
Cursor, Factory Droid, Pi	.cursor/mcp.json, .factory/mcp.json, .pi/mcp.json
OpenCode, Kilo Code	mcp in opencode.json / kilo.json

Each copy is merged into the agent’s file beside your own servers, and taken back out on uninstall.

Claude Code has no MCP file of its own at user scope (~/.claude.json is its state file), so ~/.agents/mcp.json does not reach it.

Subagents

Where they go. .agents/agents/<name>.md (~/.agents/agents at user scope), in Claude Code’s format: frontmatter name and description, then the prompt.

.agents/agents/reviewer.md

---
name: reviewer
description: Reviews a change.
---
Review the diff for bugs.

No agent reads .agents/agents yet, so each gets a copy in its own folder (.claude/agents, .cursor/agents, .opencode/agents, …), with every frontmatter key kept. Codex gets .codex/agents/<name>.toml. Pi has no subagents.

Run install again after you add or change an MCP server or a subagent.

As a tool

Command

[[mcp_servers]]
name = "mytool"
command = "mytool"
args = ["mcp"]

[[agents]]
name = "mytool-fixer"
description = "Fixes what mytool reports."
prompt = "Run `mytool check` and fix each finding."

Rust

Integration::new()
    .mcp_server(McpServer::stdio("mytool", "mytool", ["mcp"]))
    .agent(Agent::new("mytool-fixer", "Fixes what mytool reports.", "Run `mytool check`."))

Node

{
  mcp_servers: [{ name: "mytool", command: "mytool", args: ["mcp"] }],
  agents: [{ name: "mytool-fixer", description: "Fixes what mytool reports.", prompt: "Run `mytool check`." }],
}

Python

{
    "mcp_servers": [{"name": "mytool", "command": "mytool", "args": ["mcp"]}],
    "agents": [{"name": "mytool-fixer", "description": "Fixes what mytool reports.", "prompt": "Run `mytool check`."}],
}

An HTTP server takes url (and headers) in place of command.