Amp includes AGENTS.md files from the working directory and its parents, subtree AGENTS.md files when it reads files there, and personal ~/.config/amp/AGENTS.md / ~/.config/AGENTS.md. If a directory has no AGENTS.md, Amp uses AGENT.md or CLAUDE.md. Settings live in ~/.config/amp/settings.json or a workspace .amp/settings.json (amp.mcpServers, amp.mcpPermissions, amp.tools.disable). Plugins in .amp/plugins/ can allow or reject each tool call. You can sign in with ChatGPT Plus or Pro for OpenAI models.
Amp at a glance
| Global instructions | ✓ Yes~/.config/amp/AGENTS.md, ~/.config/AGENTS.md; system /etc/ampcode/AGENTS.md |
|---|---|
| Project instructions | ✓ YesAGENTS.md in cwd and parent dirs (fallback AGENT.md or CLAUDE.md) |
| Nested / subfolder | ✓ Yes Subtree AGENTS.md included when the agent reads a file in that subtree |
| AGENTS.md | ✓ Yes Primary format |
| Skills (SKILL.md) | ✓ Yes.agents/skills/ (project and parents), ~/.config/agents/skills/, ~/.agents/skills/, ~/.config/amp/skills/ |
| MCP config | ✓ Yesamp.mcpServers in ~/.config/amp/settings.json or .amp/settings.json |
| Permissions / allowlist | ◐ Partialamp.mcpPermissions, amp.tools.disable; plugins handle tool.call to allow / reject / modify |
| Sandbox | ? Not documented Local CLI sandbox not documented |
| Hooks | ✓ Yes Plugins ( .amp/plugins/, ~/.config/amp/plugins/): tool.call, tool.result, session and agent events |
| Claude Pro/Max login | ? Not documented Not documented |
| ChatGPT Plus/Pro login | ✓ Yes Sign in with ChatGPT Plus or Pro for OpenAI models |
| API key / own plan | ◐ Partial Amp credits / plans; BYOK not documented |
Exact file paths
| Path | Scope | What it does |
|---|---|---|
AGENTS.md | Project | cwd, parents and subtrees |
~/.config/amp/AGENTS.md | User | Personal guidance |
~/.config/amp/settings.json | User | amp.* settings incl. MCP |
.amp/settings.json | Workspace | Workspace settings |
.amp/plugins/ | Project | Plugins (hooks) |
.agents/skills/<name>/SKILL.md | Project | Skills |
Using a shared AGENTS.md with Amp
Native (cwd, parents and subtrees). See one AGENTS.md for every agent, or generate one with the AGENTS.md generator.
Copy-paste examples
MCP server
File: .amp/settings.json
{
"amp.mcpServers": {
"playwright": {
"command": "npx",
"args": ["-y", "@playwright/mcp@latest"]
}
}
}
Safe permissions preset
Plugin that rejects destructive shell commands
File: .amp/plugins/guard.ts
import type { PluginAPI } from '@ampcode/plugin'
export default function (amp: PluginAPI) {
amp.on('tool.call', async (event) => {
const input = JSON.stringify((event as any).input ?? '')
if (/rm -[a-z]*r[a-z]*f|git push.*(--force|-f\b)|drop (table|database)|terraform destroy/i.test(input)) {
return { action: 'reject-and-continue', message: `Blocked by project policy: ${event.tool}` }
}
return { action: 'allow' }
})
}
Built on the documented tool.call event and its allow / reject-and-continue actions. The docs don't list the field that carries the tool's arguments, so event.input is an assumption: log event once to confirm before relying on this guard.
Gotchas
- Amp removes features it doesn't use, so older guides may mention settings that no longer exist; check the current settings list.
- Project plugins run code on your machine; review
.amp/plugins/in repos you clone.
FAQ
Does Amp read CLAUDE.md?
Where does Amp store MCP servers?
Sources
Checked against these official pages on 2026-10-12. Cells marked "Not documented" mean we could not find the behaviour in these docs; it may still exist.