AI coding agent config & compatibility hub
Which instruction files, skills folders, MCP config, permission rules, sandbox and subscriptions each of 17 coding agents supports. Every cell checked against official docs on 2026-10-12.
Short answer: 12 of 17 agents read AGENTS.md natively (Codex CLI, Cursor, GitHub Copilot, Devin Desktop (ex-Windsurf), Cline, Roo Code, Kilo Code, OpenCode, Amp, Kiro, Jules, Factory Droid). Claude Code reads it only when no CLAUDE.md exists, Gemini CLI needs context.fileName, Zed lets .rules win, and Aider needs --read. Write one AGENTS.md, then add the small per-agent bridges in the one-file guide.
Free tool
AGENTS.md generator
Fill in your stack and commands; get AGENTS.md, a CLAUDE.md import line and a Cursor rule stub to copy or download.
Free tool
AGENTS.md / CLAUDE.md linter
Paste your file: size and token estimate, missing build/test commands, vague rules, leaked secrets, conflicts, absolute paths.
Guide
One AGENTS.md for every agent
AGENTS.md vs CLAUDE.md, symlinks vs imports, and the per-agent caveats that make one file work everywhere.
Compatibility matrix
Must have
| Agent | Global instructions | Project instructions | Nested / subfolder | AGENTS.md | Skills (SKILL.md) | MCP config | Permissions / allowlist | Sandbox | Hooks | Claude Pro/Max login | ChatGPT Plus/Pro login | API key / own plan | Verified |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Claude Code CLI · IDE / editor | ✓ Yes~/.claude/CLAUDE.md, ~/.claude/rules/*.md; org-managed /etc/claude-code/CLAUDE.md (Linux) or /Library/Application Support/ClaudeCode/CLAUDE.md (macOS) | ✓ YesCLAUDE.md or .claude/CLAUDE.md, personal CLAUDE.local.md, .claude/rules/*.md (optional paths: globs); loaded from the working dir and every parent dir | ✓ Yes Subdirectory CLAUDE.md loads on demand when Claude reads or edits a file there | ◐ Partial Yes since v2.1.277, but only if there is no CLAUDE.md / CLAUDE.local.md in the working dir or above (default). Set Project instructions to claude-md-and-agents-md to read both. @AGENTS.md import always works | ✓ Yes~/.claude/skills/<name>/SKILL.md, .claude/skills/<name>/SKILL.md. .agents/skills is not listed in the docs | ✓ Yes.mcp.json at project root (shared, needs approval); user and local scopes in ~/.claude.json; claude mcp add --scope project|user|local | ✓ Yespermissions.allow / ask / deny in ~/.claude/settings.json, .claude/settings.json, .claude/settings.local.json, managed settings. Deny > ask > allow. Modes: default, acceptEdits, plan, auto, dontAsk, bypassPermissions | ✓ Yes Opt-in: /sandbox or sandbox.enabled: true. Seatbelt on macOS, bubblewrap + socat on Linux/WSL2 | ✓ Yeshooks in settings.json: PreToolUse, PostToolUse, SessionStart, Stop, UserPromptSubmit and more; exit code 2 blocks a tool call | ✓ Yes Log in with Claude Pro or Max (also Team/Enterprise) | ✕ No Not supported | ✓ YesANTHROPIC_API_KEY, Claude Console, Bedrock, Google Cloud, Microsoft Foundry | 2026-10-12 source |
| Codex CLI CLI · IDE / editor · Cloud | ✓ Yes~/.codex/AGENTS.override.md or ~/.codex/AGENTS.md ($CODEX_HOME); first non-empty file wins | ✓ Yes In each dir from git root to cwd: AGENTS.override.md, else AGENTS.md, else names in project_doc_fallback_filenames; max one per dir; 32 KiB total (project_doc_max_bytes) | ◐ Partial Directories between the repo root and where you launch Codex; files below the launch dir are not read | ✓ Yes Native, primary format. CLAUDE.md only if you add it to project_doc_fallback_filenames | ✓ Yes.agents/skills from cwd up to repo root, ~/.agents/skills, /etc/codex/skills | ✓ Yes[mcp_servers.<name>] in ~/.codex/config.toml or project .codex/config.toml (trusted projects only); codex mcp add | ✓ Yesapproval_policy = on-request | never | granular (untrusted retired, on-failure deprecated); .rules files with prefix_rule() decisions allow / prompt / forbidden | ✓ Yes On by default: sandbox_mode read-only / workspace-write / danger-full-access; no network by default. Seatbelt (macOS), bubblewrap (Linux/WSL2), native Windows sandbox | ✓ Yes~/.codex/hooks.json, <repo>/.codex/hooks.json or [hooks] in config.toml; PreToolUse can deny; enabled by default | ✕ No Not supported | ✓ Yes Sign in with ChatGPT uses your plan's Codex usage (required for Codex cloud) | ✓ Yes Sign in with an API key for usage-based billing | 2026-10-12 source |
| Cursor IDE / editor · CLI · Cloud | ✓ Yes User Rules in Customize → Rules (stored in Cursor, not a file); Team Rules from the dashboard | ✓ Yes.cursor/rules/*.mdc (frontmatter: description, globs, alwaysApply), AGENTS.md. Precedence Team → Project → User | ✓ Yes Nested AGENTS.md in any subdirectory; .cursor/rules/ subfolders | ✓ Yes Root and nested AGENTS.md, combined with parents | ✓ Yes.cursor/skills/, .agents/skills/, ~/.cursor/skills/, ~/.agents/skills/; also .claude/skills/ and .codex/skills/ for compatibility | ✓ Yes.cursor/mcp.json (project) and ~/.cursor/mcp.json (global), mcpServers object | ✓ Yes IDE Run Modes: Auto-review, Allowlist, Run Everything; permissions.json steers Auto-review. CLI: permissions.allow/deny with Shell(), Read(), Write(), WebFetch() in ~/.cursor/cli-config.json or .cursor/cli.json | ✓ Yes Terminal sandbox; configure with ~/.cursor/sandbox.json or .cursor/sandbox.json; network blocked by default inside it | ✓ Yes.cursor/hooks.json and ~/.cursor/hooks.json: preToolUse, beforeShellExecution, afterFileEdit and more; can load Claude Code hooks | ✕ No Not a login option; add an Anthropic API key instead | ✕ No Not a login option; add an OpenAI API key instead | ✓ Yes BYOK for OpenAI, Anthropic, Google, Azure OpenAI, AWS Bedrock (chat models only; Tab stays on Cursor models) | 2026-10-12 source |
| GitHub Copilot IDE / editor · Cloud · CLI | ✓ Yes Personal and organization instructions on GitHub.com; ~/.copilot/copilot-instructions.md for Copilot Agent Host / CLI sessions | ✓ Yes.github/copilot-instructions.md, .github/instructions/**/*.instructions.md (applyTo), AGENTS.md; cloud agent also CLAUDE.md or GEMINI.md at root | ◐ Partial Cloud agent: AGENTS.md anywhere, nearest wins. VS Code: nested AGENTS.md is experimental ( chat.useNestedAgentsMdFiles, off by default) | ✓ Yes VS Code agent mode and cloud agent; toggle chat.useAgentsMdFile in VS Code | ✓ Yes.github/skills, .claude/skills, .agents/skills; personal ~/.copilot/skills, ~/.agents/skills (cloud agent, CLI, VS Code) | ✓ Yes VS Code: .vscode/mcp.json (servers) or .mcp.json (mcpServers), plus user-profile mcp.json. Cloud agent: JSON in repository Settings (mcpServers) | ✓ Yes VS Code permission levels and Autopilot; chat.tools.terminal.autoApprove (true / false / regex), chat.tools.edits.autoApprove, chat.tools.urls.autoApprove | ✓ Yes VS Code agent sandboxing (Preview on macOS/Linux/WSL2, experimental on Windows). Cloud agent: ephemeral cloud environment + internet firewall | ✓ Yes.github/hooks/*.json for cloud agent and Copilot CLI (personal ~/.copilot/hooks/*.json); VS Code also reads .github/hooks/ (Preview) | ✕ No Not documented | ✕ No Not documented | ◐ Partial Copilot plan required; bring-your-own-key options not verified here | 2026-10-12 source |
| Antigravity / Gemini CLI CLI · IDE / editor | ✓ Yes Antigravity: ~/.gemini/AGENTS.md, ~/.gemini/GEMINI.md, ~/.gemini/config/AGENTS.md. Gemini CLI: ~/.gemini/GEMINI.md | ✓ Yes Antigravity: AGENTS.md / GEMINI.md, .agents/AGENTS.md, .agents/rules/*.md. Gemini CLI: GEMINI.md in workspace dirs and parents | ✓ Yes Directory-scoped rules at each level; Gemini CLI loads GEMINI.md just-in-time when a tool touches a directory | ◐ Partial Antigravity: yes, native. Gemini CLI: only via context.fileName (e.g. ["AGENTS.md", "GEMINI.md"]) | ✓ Yes.agents/skills/ (workspace), ~/.gemini/config/skills/ (Antigravity), ~/.gemini/antigravity-cli/skills/ (CLI); Gemini CLI: .gemini/skills/ or .agents/skills/, ~/.gemini/skills/ | ✓ Yes Antigravity: ~/.gemini/config/mcp_config.json, .agents/mcp_config.json. Gemini CLI: mcpServers in ~/.gemini/settings.json or .gemini/settings.json | ✓ Yes Antigravity: Deny > Ask > Allow rules ( command(), read_file(), write_file(), read_url(), mcp()) over a Default / Request Review / Turbo preset. Gemini CLI: policy engine TOML in ~/.gemini/policies/ (allow / deny / ask_user) | ✓ Yes Antigravity terminal sandbox on under the Default preset (workspace + temp only, no network). Gemini CLI: --sandbox / -s (Docker, Podman or macOS Seatbelt) | ✓ Yes Antigravity: .agents/hooks.json, ~/.gemini/config/hooks.json. Gemini CLI: hooks in settings.json | ✕ No Not supported | ✕ No Not supported | ✓ Yes Google account sign-in (free, Google AI Pro/Ultra, enterprise) or a Gemini API key; Vertex AI for Gemini CLI | 2026-10-12 source |
| Devin Desktop (ex-Windsurf) Changed IDE / editor · CLI | ✓ Yes Cascade: ~/.codeium/windsurf/memories/global_rules.md (6,000 chars). Devin Local/CLI: ~/.config/devin/AGENTS.md, ~/.claude/CLAUDE.md | ✓ Yes.devin/rules/*.md (preferred) or .windsurf/rules/*.md, 12,000 chars per file; AGENTS.md; Devin Local also AGENTS.local.md, AGENT.md, CLAUDE.md, .windsurfrules, .cursor/rules/*.mdc | ✓ Yes Subdirectory AGENTS.md is scoped to that folder; rules dirs found in subfolders and parents up to the git root | ✓ Yes Root AGENTS.md is an always-on rule; nested files apply to their directory | ✓ Yes.devin/skills/ or .windsurf/skills/, .agents/skills/; global ~/.config/devin/skills/, ~/.agents/skills/ | ✓ Yes Devin Local/CLI: .devin/mcp_config.json (project), .devin/mcp_config.local.json (local), ~/.config/devin/mcp_config.json (user). Legacy Cascade: its own mcp_config.json | ✓ Yes Devin Local: deny / ask / allow rules ( Read(), Write(), Exec(), fetch, MCP) in .devin/config.json or ~/.config/devin/config.json | ✓ Yes Devin Local supports OS-level sandboxing (filesystem + network filtering) | ✓ Yes Devin CLI/Local hooks in .devin/; Cascade Hooks for the legacy agent | ✕ No Not documented | ✕ No Not documented | ? Not documented Uses Devin/Windsurf plan quota; BYOK not documented in the pages checked | 2026-10-12 source |
| Cline IDE / editor · CLI | ✓ Yes~/Documents/Cline/Rules (also ~/.cline/rules), ~/.agents/AGENTS.md | ✓ Yes.clinerules/ or .cline/rules/; auto-detects .cursorrules, .windsurfrules, AGENTS.md | ? Not documented Not documented | ✓ YesAGENTS.md and global ~/.agents/AGENTS.md | ✓ Yes.cline/skills/ (recommended), .clinerules/skills/, .claude/skills/; global ~/.cline/skills/ | ✓ Yes IDE: MCP Servers → Configure → cline_mcp_settings.json (mcpServers). CLI: ~/.cline/data/settings/cline_mcp_settings.json | ◐ Partial Auto Approve per category (read, edit, safe commands, all commands) and YOLO mode; the model flags commands as requiring approval, no fixed allowlist | ? Not documented Not documented for the local extension | ◐ Partial Lifecycle hooks via plugins | ✓ Yes Claude Code provider: uses your Claude Pro/Max limits via the installed, logged-in Claude CLI | ✓ Yes OpenAI Codex provider: sign in with your ChatGPT account (OAuth) | ✓ Yes Anthropic, OpenAI, OpenRouter, Bedrock, Gemini and many more providers | 2026-10-12 source |
| Roo Code Discontinued IDE / editor | ✓ Yes~/.roo/rules/, ~/.roo/rules-<mode>/ | ✓ Yes.roo/rules/ (preferred) or .roorules; .roo/rules-<mode>/ | ? Not documented Not documented | ✓ Yes Root AGENTS.md (or AGENT.md), on by default | ✓ Yes.roo/skills/, .agents/skills/; global ~/.roo/skills/, ~/.agents/skills/ | ✓ Yes Global mcp_settings.json; project .roo/mcp.json | ✓ Yes Auto-approve with allowed / denied command prefix lists | ? Not documented Not documented | ? Not documented Not documented | ? Not documented Not verified (product discontinued) | ? Not documented Not verified (product discontinued) | ✓ Yes Your own provider keys | 2026-10-12 source |
| Kilo Code IDE / editor · CLI | ✓ Yesinstructions in ~/.config/kilo/kilo.jsonc; ~/.kilo/rules/*.md | ✓ YesAGENTS.md / AGENT.md; files under instructions in kilo.jsonc or .kilo/kilo.jsonc (e.g. .kilo/rules/*.md); legacy .kilocode/rules/ | ✓ Yes Subdirectory AGENTS.md loads when the agent accesses files there | ✓ Yes Loaded automatically; replaces the deprecated memory bank | ✓ Yes.kilo/skills/, ~/.kilo/skills/, .agents/skills/, ~/.agents/skills/; .claude/skills/ with Claude Code compatibility on | ✓ Yesmcp key in kilo.jsonc / .kilo/kilo.jsonc or ~/.config/kilo/kilo.jsonc | ✓ Yespermission in kilo.jsonc: allow / ask / deny per tool (read, edit, bash, webfetch…) with patterns | ✓ Yes Optional OS sandbox (off by default; macOS/Linux, not Windows): limits writes and blocks network | ◐ Partial Plugins ( .kilo/plugin/*.ts) with events such as tool.execute.before | ? Not documented Not documented | ✓ Yes Connect a ChatGPT Plus/Pro subscription for supported OpenAI models | ✓ Yes Kilo Gateway credits or your own provider keys | 2026-10-12 source |
| Aider CLI | ◐ Partialread: in ~/.aider.conf.yml | ◐ Partial Only files you pass with --read / /read or list under read: in .aider.conf.yml | ✕ No No automatic discovery | ◐ Partial Not auto-loaded; read: AGENTS.md works like any conventions file | ? Not documented Not documented | ? Not documented Not documented in Aider's docs | ◐ Partial Confirms suggested shell commands; --yes-always skips confirmations; auto-commits every edit | ? Not documented Not documented | ◐ Partial--lint-cmd / --test-cmd run after edits (no general hook system documented) | ? Not documented Not documented | ? Not documented Not documented | ✓ Yes API keys for Anthropic, OpenAI, Gemini, DeepSeek, OpenRouter, local models and more | 2026-10-12 source |
| OpenCode CLI · IDE / editor | ✓ Yes~/.config/opencode/AGENTS.md, else ~/.claude/CLAUDE.md | ✓ YesAGENTS.md (else CLAUDE.md) found by traversing up from cwd; instructions array in opencode.json | ◐ Partial Parent directories by traversing up; first match wins | ✓ Yes Primary format | ✓ Yes.opencode/skills/, .claude/skills/, .agents/skills/; global ~/.config/opencode/skills/, ~/.claude/skills/, ~/.agents/skills/ | ✓ Yesmcp key in opencode.json / opencode.jsonc (project or ~/.config/opencode/) | ✓ Yespermission in opencode.json: allow / ask / deny per tool (bash, edit, webfetch, external_directory…) with wildcards; last matching rule wins | ? Not documented Not documented | ✓ Yes Plugins in .opencode/plugins/ subscribe to events such as tool execution | ✕ No Not supported: Anthropic prohibits Claude Pro/Max in third-party tools; OpenCode removed those plugins in v1.3.0 | ✓ Yes ChatGPT Plus works with zero setup (also GitHub Copilot, GitLab Duo) | ✓ Yes 75+ providers via API keys, local models | 2026-10-12 source |
| Zed IDE / editor | ✓ Yes~/.config/zed/AGENTS.md | ✓ Yes First match of .rules, .cursorrules, .windsurfrules, .clinerules, .github/copilot-instructions.md, AGENT.md, AGENTS.md, CLAUDE.md, GEMINI.md | ? Not documented Not documented | ◐ Partial Yes, but .rules, .cursorrules, .windsurfrules, .clinerules, copilot-instructions.md or AGENT.md take priority | ✓ Yes~/.agents/skills/ and <worktree>/.agents/skills/ (flat layout) | ✓ Yescontext_servers in Zed settings.json, or MCP extensions | ✓ Yesagent.tool_permissions (v0.224+): default plus per-tool always_allow / always_confirm / always_deny regex patterns | ? Not documented Not documented | ? Not documented Not documented | ◐ Partial Only via the Claude Agent (ACP), which has its own login | ◐ Partial Only via the Codex external agent | ✓ Yes Zed's hosted models or your own provider keys | 2026-10-12 source |
| Amp CLI · Cloud | ✓ Yes~/.config/amp/AGENTS.md, ~/.config/AGENTS.md; system /etc/ampcode/AGENTS.md | ✓ YesAGENTS.md in cwd and parent dirs (fallback AGENT.md or CLAUDE.md) | ✓ Yes Subtree AGENTS.md included when the agent reads a file in that subtree | ✓ Yes Primary format | ✓ Yes.agents/skills/ (project and parents), ~/.config/agents/skills/, ~/.agents/skills/, ~/.config/amp/skills/ | ✓ Yesamp.mcpServers in ~/.config/amp/settings.json or .amp/settings.json | ◐ Partialamp.mcpPermissions, amp.tools.disable; plugins handle tool.call to allow / reject / modify | ? Not documented Local CLI sandbox not documented | ✓ Yes Plugins ( .amp/plugins/, ~/.config/amp/plugins/): tool.call, tool.result, session and agent events | ? Not documented Not documented | ✓ Yes Sign in with ChatGPT Plus or Pro for OpenAI models | ◐ Partial Amp credits / plans; BYOK not documented | 2026-10-12 source |
| Kiro IDE / editor · CLI · Cloud | ✓ Yes~/.kiro/steering/*.md (AGENTS.md allowed there) | ✓ Yes.kiro/steering/*.md with inclusion modes; AGENTS.md at root | ? Not documented Not documented | ✓ Yes Always included (no inclusion modes) | ✓ Yes.kiro/skills/ (workspace), ~/.kiro/skills/ (global) | ✓ Yes.kiro/settings/mcp.json (workspace), ~/.kiro/settings/mcp.json (user) | ✓ Yes Capability rules ( shell, fs_read, fs_write, web_fetch, mcp) with deny > ask > allow in ~/.kiro/settings/permissions.yaml | ◐ Partial Kiro on the web runs in an isolated cloud sandbox; local sandbox not detailed | ✓ Yes.kiro/hooks/ (file events, tool events) | ✕ No Not supported | ✕ No Not supported | ✕ No Kiro plans; sign in with GitHub, Google, AWS Builder ID or IAM Identity Center | 2026-10-12 source |
| Jules Cloud | ? Not documented Not documented | ✓ YesAGENTS.md in the repository root | ? Not documented Not documented | ✓ Yes Root AGENTS.md, read automatically | ? Not documented Not documented | ◐ Partial Vetted servers only (e.g. Linear, Neon, Supabase, Context7), configured in Settings with API keys | ◐ Partial Plan review and approval before coding | ✓ Yes Every task runs in a fresh short-lived VM | ? Not documented Not documented | ✕ No Not supported | ✕ No Not supported | ✕ No Google account; higher limits via Google AI Pro / Ultra | 2026-10-12 source |
| Factory Droid CLI · IDE / editor · Cloud | ✓ Yes~/.factory/, ~/.agents/, ~/.agent/ (AGENTS.md) | ✓ YesAGENTS.md in each dir from cwd up to git root, and in .factory/, .agents/, .agent/ | ✓ Yes Nested AGENTS.md refines root files | ✓ Yes Primary format | ✓ Yes.factory/skills/, .agents/skills/; ~/.factory/skills/, ~/.agents/skills/ | ✓ Yes~/.factory/mcp.json (user), .factory/mcp.json (project) | ✓ YespermissionRules (allow / ask / block, token-prefix match, required tests) in ~/.factory/settings.json or .factory/settings.json; autonomy levels | ✓ Yes OS sandbox: Seatbelt (macOS), bubblewrap (Linux/WSL2); per-command or whole-process | ✓ Yes~/.factory/hooks.json, .factory/hooks.json | ? Not documented Not documented | ? Not documented Not documented | ✓ Yes Factory plans or BYOK custom models | 2026-10-12 source |
| Continue Discontinued IDE / editor · CLI | ? Not documented Not documented on the rules page | ✓ Yes.continue/rules/*.md (Markdown or YAML) | ? Not documented Not documented | ? Not documented Not documented | ? Not documented Not documented | ✓ YesmcpServers in config, or files in .continue/mcpServers/ (accepts Claude/Cursor/Cline JSON) | ? Not documented Not documented on the pages checked | ? Not documented Not documented | ? Not documented Not documented | ? Not documented Not documented | ? Not documented Not documented | ✓ Yes Your own keys or local models | 2026-10-12 source |
Legend: ✓ Yes ◐ Partial / with caveats ✕ No ? Not documented in the vendor docs we checked. Raw data: config.json. Found an error? Each agent page links the exact doc pages we used.
Instruction files by agent
Which project file each agent reads, and whether it accepts AGENTS.md:
| Agent | Project instruction files | AGENTS.md |
|---|---|---|
| Claude Code | CLAUDE.md or .claude/CLAUDE.md, personal CLAUDE.local.md, .claude/rules/*.md (optional paths: globs); loaded from the working dir and every parent dir | ◐ Partial |
| Codex CLI | In each dir from git root to cwd: AGENTS.override.md, else AGENTS.md, else names in project_doc_fallback_filenames; max one per dir; 32 KiB total (project_doc_max_bytes) | ✓ Yes |
| Cursor | .cursor/rules/*.mdc (frontmatter: description, globs, alwaysApply), AGENTS.md. Precedence Team → Project → User | ✓ Yes |
| GitHub Copilot | .github/copilot-instructions.md, .github/instructions/**/*.instructions.md (applyTo), AGENTS.md; cloud agent also CLAUDE.md or GEMINI.md at root | ✓ Yes |
| Antigravity / Gemini CLI | Antigravity: AGENTS.md / GEMINI.md, .agents/AGENTS.md, .agents/rules/*.md. Gemini CLI: GEMINI.md in workspace dirs and parents | ◐ Partial |
| Devin Desktop (ex-Windsurf) | .devin/rules/*.md (preferred) or .windsurf/rules/*.md, 12,000 chars per file; AGENTS.md; Devin Local also AGENTS.local.md, AGENT.md, CLAUDE.md, .windsurfrules, .cursor/rules/*.mdc | ✓ Yes |
| Cline | .clinerules/ or .cline/rules/; auto-detects .cursorrules, .windsurfrules, AGENTS.md | ✓ Yes |
| Kilo Code | AGENTS.md / AGENT.md; files under instructions in kilo.jsonc or .kilo/kilo.jsonc (e.g. .kilo/rules/*.md); legacy .kilocode/rules/ | ✓ Yes |
| Aider | Only files you pass with --read / /read or list under read: in .aider.conf.yml | ◐ Partial |
| OpenCode | AGENTS.md (else CLAUDE.md) found by traversing up from cwd; instructions array in opencode.json | ✓ Yes |
| Zed | First match of .rules, .cursorrules, .windsurfrules, .clinerules, .github/copilot-instructions.md, AGENT.md, AGENTS.md, CLAUDE.md, GEMINI.md | ◐ Partial |
| Amp | AGENTS.md in cwd and parent dirs (fallback AGENT.md or CLAUDE.md) | ✓ Yes |
| Kiro | .kiro/steering/*.md with inclusion modes; AGENTS.md at root | ✓ Yes |
| Jules | AGENTS.md in the repository root | ✓ Yes |
| Factory Droid | AGENTS.md in each dir from cwd up to git root, and in .factory/, .agents/, .agent/ | ✓ Yes |
MCP config file location by agent
No two agents share an MCP file, though most use the same mcpServers JSON shape (VS Code uses servers; Codex uses TOML; OpenCode and Kilo use mcp).
| Agent | MCP config |
|---|---|
| Claude Code | .mcp.json at project root (shared, needs approval); user and local scopes in ~/.claude.json; claude mcp add --scope project|user|local |
| Codex CLI | [mcp_servers.<name>] in ~/.codex/config.toml or project .codex/config.toml (trusted projects only); codex mcp add |
| Cursor | .cursor/mcp.json (project) and ~/.cursor/mcp.json (global), mcpServers object |
| GitHub Copilot | VS Code: .vscode/mcp.json (servers) or .mcp.json (mcpServers), plus user-profile mcp.json. Cloud agent: JSON in repository Settings (mcpServers) |
| Antigravity / Gemini CLI | Antigravity: ~/.gemini/config/mcp_config.json, .agents/mcp_config.json. Gemini CLI: mcpServers in ~/.gemini/settings.json or .gemini/settings.json |
| Devin Desktop (ex-Windsurf) | Devin Local/CLI: .devin/mcp_config.json (project), .devin/mcp_config.local.json (local), ~/.config/devin/mcp_config.json (user). Legacy Cascade: its own mcp_config.json |
| Cline | IDE: MCP Servers → Configure → cline_mcp_settings.json (mcpServers). CLI: ~/.cline/data/settings/cline_mcp_settings.json |
| Kilo Code | mcp key in kilo.jsonc / .kilo/kilo.jsonc or ~/.config/kilo/kilo.jsonc |
| Aider | Not documented in Aider's docs |
| OpenCode | mcp key in opencode.json / opencode.jsonc (project or ~/.config/opencode/) |
| Zed | context_servers in Zed settings.json, or MCP extensions |
| Amp | amp.mcpServers in ~/.config/amp/settings.json or .amp/settings.json |
| Kiro | .kiro/settings/mcp.json (workspace), ~/.kiro/settings/mcp.json (user) |
| Jules | Vetted servers only (e.g. Linear, Neon, Supabase, Context7), configured in Settings with API keys |
| Factory Droid | ~/.factory/mcp.json (user), .factory/mcp.json (project) |
Claude / ChatGPT subscription vs API key
| Agent | Claude Pro/Max | ChatGPT Plus/Pro | API key / own plan |
|---|---|---|---|
| Claude Code | ✓ Yes Log in with Claude Pro or Max (also Team/Enterprise) | ✕ No Not supported | ANTHROPIC_API_KEY, Claude Console, Bedrock, Google Cloud, Microsoft Foundry |
| Codex CLI | ✕ No Not supported | ✓ Yes Sign in with ChatGPT uses your plan's Codex usage (required for Codex cloud) | Sign in with an API key for usage-based billing |
| Cursor | ✕ No Not a login option; add an Anthropic API key instead | ✕ No Not a login option; add an OpenAI API key instead | BYOK for OpenAI, Anthropic, Google, Azure OpenAI, AWS Bedrock (chat models only; Tab stays on Cursor models) |
| GitHub Copilot | ✕ No Not documented | ✕ No Not documented | Copilot plan required; bring-your-own-key options not verified here |
| Antigravity / Gemini CLI | ✕ No Not supported | ✕ No Not supported | Google account sign-in (free, Google AI Pro/Ultra, enterprise) or a Gemini API key; Vertex AI for Gemini CLI |
| Devin Desktop (ex-Windsurf) | ✕ No Not documented | ✕ No Not documented | Uses Devin/Windsurf plan quota; BYOK not documented in the pages checked |
| Cline | ✓ Yes Claude Code provider: uses your Claude Pro/Max limits via the installed, logged-in Claude CLI | ✓ Yes OpenAI Codex provider: sign in with your ChatGPT account (OAuth) | Anthropic, OpenAI, OpenRouter, Bedrock, Gemini and many more providers |
| Kilo Code | ? Not documented Not documented | ✓ Yes Connect a ChatGPT Plus/Pro subscription for supported OpenAI models | Kilo Gateway credits or your own provider keys |
| Aider | ? Not documented Not documented | ? Not documented Not documented | API keys for Anthropic, OpenAI, Gemini, DeepSeek, OpenRouter, local models and more |
| OpenCode | ✕ No Not supported: Anthropic prohibits Claude Pro/Max in third-party tools; OpenCode removed those plugins in v1.3.0 | ✓ Yes ChatGPT Plus works with zero setup (also GitHub Copilot, GitLab Duo) | 75+ providers via API keys, local models |
| Zed | ◐ Partial Only via the Claude Agent (ACP), which has its own login | ◐ Partial Only via the Codex external agent | Zed's hosted models or your own provider keys |
| Amp | ? Not documented Not documented | ✓ Yes Sign in with ChatGPT Plus or Pro for OpenAI models | Amp credits / plans; BYOK not documented |
| Kiro | ✕ No Not supported | ✕ No Not supported | Kiro plans; sign in with GitHub, Google, AWS Builder ID or IAM Identity Center |
| Jules | ✕ No Not supported | ✕ No Not supported | Google account; higher limits via Google AI Pro / Ultra |
| Factory Droid | ? Not documented Not documented | ? Not documented Not documented | Factory plans or BYOK custom models |
Safe-permission presets
Every agent page ends with a copy-paste preset that blocks rm -rf, git push --force, git reset --hard and DROP TABLE, keeps .env and ~/.ssh out of reach and asks before git push. Command rules match text, so turn on the sandbox too.
- Claude Code: Safe settings.json: block destructive commands and secrets
- Codex CLI: Safe Codex setup: workspace sandbox, no network, forbidden commands
- Cursor: Safe Cursor CLI permissions (deny destructive commands and secrets)
- GitHub Copilot: Safe VS Code settings for Copilot agent mode
- Antigravity / Gemini CLI: Safe Antigravity permissions (Deny / Ask lists)
- Devin Desktop (ex-Windsurf): Safe Devin Local permissions
- Cline: Safe Cline Auto Approve settings
- Kilo Code: Safe Kilo permissions
- Aider: Safe Aider config
- OpenCode: Safe opencode.json permissions
- Zed: Safe Zed tool permissions
- Amp: Plugin that rejects destructive shell commands
- Kiro: Safe Kiro permissions
- Jules: Safe Jules workflow
- Factory Droid: Safe Droid permission rules
FAQ
Which AI coding agents support AGENTS.md?
As of 2026-10-12, these agents read AGENTS.md natively: Codex CLI, Cursor, GitHub Copilot, Devin Desktop (ex-Windsurf), Cline, Roo Code, Kilo Code, OpenCode, Amp, Kiro, Jules, Factory Droid. With caveats: Claude Code — Yes since v2.1.277, but only if there is no CLAUDE.md / CLAUDE.local.md in the working dir or above (default). Set Project instructions to claude-md-and-agents-md to read both. @AGENTS.md import always works; Antigravity / Gemini CLI — Antigravity: yes, native. Gemini CLI: only via context.fileName (e.g. ["AGENTS.md", "GEMINI.md"]); Aider — Not auto-loaded; read: AGENTS.md works like any conventions file; Zed — Yes, but .rules, .cursorrules, .windsurfrules, .clinerules, copilot-instructions.md or AGENT.md take priority.
AGENTS.md vs CLAUDE.md: which should I use?
Use AGENTS.md as the single source of truth because most agents read it. Claude Code prefers CLAUDE.md: by default it reads AGENTS.md only when no CLAUDE.md exists, so add a CLAUDE.md whose first line is @AGENTS.md (or a symlink) and put Claude-only notes below it.
Where is the MCP config file for each AI coding agent?
There is no shared location. Claude Code uses .mcp.json, Cursor .cursor/mcp.json, VS Code Copilot .vscode/mcp.json, Codex ~/.codex/config.toml, Antigravity ~/.gemini/config/mcp_config.json, Kiro .kiro/settings/mcp.json, OpenCode the mcp key in opencode.json. The matrix lists every agent.
Which coding agents can use my Claude Pro or Max subscription?
Officially: Claude Code, Cline. OpenCode explicitly does not support it because Anthropic prohibits Claude subscriptions in third-party tools; Cursor and Copilot bill through their own plans.
Which coding agents can use a ChatGPT Plus or Pro subscription?
Codex CLI, Cline, Kilo Code, OpenCode, Amp support signing in with ChatGPT. Most other agents need an OpenAI API key.
What is the safest permission setup for an AI coding agent?
Deny destructive commands (rm -rf, git push --force, git reset --hard, DROP TABLE, terraform destroy) and reads of .env and ~/.ssh, require approval for git push, and turn on the agent's sandbox. Command rules match text, so pair them with a sandbox or a pre-tool hook. Each agent page has a copy-paste preset.