Skip to content
English

Switch theme

Discover▾

Not sure where to start? This week’s hottest software, models and stacks in three minutes.

AI Apps▾

AI software people use: from assistants like ChatGPT to tools for coding, video and workflows, picked by purpose.

For Agents▾

Software for AI: tools that let your agent (Claude Code, Codex, Cursor…) browse, read docs, run code and remember.

Models▾

Choose models, estimate cost, self-host: prices, rankings, benchmarks, free tiers and compute.

Solutions▾

A goal or an error message? Ready-made stacks, steps and fixes by scenario.

Learn▾

A systematic course from zero to the frontier, with an algorithm library, must-read papers and tutorials.

Frontier▾

Daily AI news, weekly highlights, the latest papers and how AI is changing every discipline.

// Agent tools · setup guide

Agent setup guide

Installing an MCP server, writing rules or adding Skills? Where each coding agent keeps its config, and how much it lets the agent do on its own — one table, official sources only.

Open the table ↓ Safety checklist →

11 agentschecked 2026-10-22MCP spec 2026-07-28official docs only

01Config file cheat sheet

~ is your home directory. “Project” files can be committed so the whole team gets them.

AgentMCP configAdd from terminalTransportsRules / instructionsSkills folders
Claude CodeAnthropic · CLI / IDE / Desktop / Web ~/.claude.json (local / user scope).mcp.json (project scope, shared) claude mcp add --transport http <name> <url> stdio · HTTP · SSE (deprecated) · WebSocketOAuth ✓ CLAUDE.md / .claude/CLAUDE.mdCLAUDE.local.md~/.claude/CLAUDE.md.claude/rules/AGENTS.md (read when present) .claude/skills/<name>/SKILL.md~/.claude/skills/
Codex (CLI · IDE · ChatGPT desktop)OpenAI · CLI / IDE / Desktop ~/.codex/config.toml → [mcp_servers.<name>].codex/config.toml (trusted projects only) codex mcp add <name> -- <command> stdio · Streamable HTTPOAuth ✓ AGENTS.md (repo root → cwd)AGENTS.override.md~/.codex/AGENTS.md .agents/skills/ (cwd → repo root)~/.agents/skills//etc/codex/skills
Cursor (editor · CLI)Anysphere · IDE / CLI .cursor/mcp.json (project)~/.cursor/mcp.json (global) UI or edit file stdio · SSE · Streamable HTTPOAuth ✓ .cursor/rules/*.mdcAGENTS.mdUser Rules / Team Rules (settings) .agents/skills/.cursor/skills/~/.agents/skills/~/.cursor/skills/.claude/skills/ · .codex/skills/ (compat)
GitHub Copilot (VS Code · Copilot CLI)GitHub · IDE / CLI / Cloud .vscode/mcp.json (VS Code)~/.copilot/mcp-config.json (CLI).mcp.json · .github/mcp.json (CLI, per repo) copilot mcp add · /mcp add stdio · HTTP .github/copilot-instructions.md.github/instructions/*.instructions.mdAGENTS.mdCLAUDE.md / GEMINI.md (repo root) .github/skills/.claude/skills/.agents/skills/~/.copilot/skills/~/.agents/skills/
Antigravity CLIGoogle · CLI / Desktop⚠ Launched 2026-05-19 as Gemini CLI’s successor; since 2026-06-18 Gemini CLI no longer serves free and Google AI Pro / Ultra individual users ~/.gemini/config/mcp_config.json (global).agents/mcp_config.json (workspace) UI or edit file stdio · SSE · Streamable HTTP · WebSocket GEMINI.mdAGENTS.md~/.gemini/GEMINI.md .agents/skills/~/.gemini/antigravity-cli/skills/
Gemini CLIGoogle · CLI⚠ Still served only for enterprise (Gemini Code Assist licences) and API-key users; individuals should move to Antigravity CLI ~/.gemini/settings.json → mcpServers.gemini/settings.json (project, default scope) gemini mcp add <name> <command|url> stdio · SSE · Streamable HTTPOAuth ✓ GEMINI.md (context.fileName configurable)~/.gemini/GEMINI.md .gemini/skills/ · .agents/skills/~/.gemini/skills/ · ~/.agents/skills/
Devin Desktop / Devin CLI (formerly Windsurf)Cognition · IDE / CLI / Cloud⚠ Windsurf became Devin Desktop on 2026-06-02; the default local agent moved from Cascade to Devin Local (shares Devin CLI config) .devin/mcp_config.local.json (local, default).devin/mcp_config.json (project)~/.config/devin/mcp_config.json (user) devin mcp add <name> <url> stdio · Streamable HTTP · SSEOAuth ✓ AGENTS.md~/.config/devin/AGENTS.md.devin/rules/ (legacy .windsurf/rules/) .agents/skills/.devin/skills/ (legacy .windsurf/skills/)~/.agents/skills/~/.config/devin/skills/
ClineCline · IDE / CLI cline_mcp_settings.json (extension: MCP Servers → Configure)~/.cline/data/settings/cline_mcp_settings.json (CLI) UI or edit file stdio · Streamable HTTP (recommended) · SSE (legacy) .clinerules/ · .cline/rules/AGENTS.md · ~/.agents/AGENTS.md.cursorrules · .windsurfrules (auto-detected) .cline/skills/ · .clinerules/skills/.claude/skills/~/.cline/skills/
OpenCodeAnomaly · CLI / IDE / Desktop opencode.json(c) → "mcp" (type: local | remote) UI or edit file local (stdio) · remote (HTTP)OAuth ✓ AGENTS.md~/.config/opencode/AGENTS.mdCLAUDE.md · ~/.claude/CLAUDE.md (fallback) .opencode/skills/~/.config/opencode/skills/.claude/skills/ · ~/.claude/skills/.agents/skills/ · ~/.agents/skills/
Kiro (IDE · CLI)AWS · IDE / CLI / Web .kiro/settings/mcp.json (workspace)~/.kiro/settings/mcp.json (user) UI or edit file stdio · remote HTTPSOAuth ✓ .kiro/steering/~/.kiro/steering/AGENTS.md .kiro/skills/~/.kiro/skills/
Qwen CodeAlibaba Qwen · CLI ~/.qwen/settings.json (user, default).qwen/settings.json (project) qwen mcp add --transport http <name> <url> stdio · Streamable HTTP · SSEOAuth ✓ QWEN.md (context.fileName configurable) .qwen/skills/~/.qwen/skills/

02Write once, use in several agents

Counts are from the table above.

  1. AGENTS.md9 of 11 read it. Put shared project rules here; keep CLAUDE.md, GEMINI.md or .cursor/rules for tool-specific differences. agents.md ↗
  2. .agents/skills/7 of 11 scan it. One folder per skill with a SKILL.md (name + description in frontmatter) — the open Agent Skills format. agentskills.io ↗
  3. mcpServers JSONMost clients accept the same mcpServers shape, but check: url entries need "type" in Claude Code, Antigravity CLI wants serverUrl, Codex uses TOML.
  4. Prefer remote (HTTP + OAuth)Every client here supports a remote transport. A remote server doesn’t run code on your machine and signs in with scoped OAuth instead of a long-lived token. Why →

03Each agent: permissions, sandbox, extensions

“not documented” means the official docs we checked don’t say — not that the feature is missing.

Claude Code Anthropic

Permissions
Permission modes: default (ask per tool) / acceptEdits / plan / auto (classifier review) / dontAsk / bypassPermissions
Sandbox
Turn on with /sandbox (off by default); Seatbelt on macOS, bubblewrap on Linux / WSL2, not on native Windows
MCP
Project .mcp.json servers need your approval before first use; /mcp shows status and runs OAuth
Extensions
Hooks ✓Subagents ✓Plugins ✓

MCP ↗CLAUDE.md / AGENTS.md ↗Skills ↗Permissions ↗Sandbox ↗

Profile on AIMap →

Codex (CLI · IDE · ChatGPT desktop) OpenAI

Permissions
Sandbox read-only / workspace-write (default, network off) / danger-full-access; approvals on-request / never / granular
Sandbox
On by default: Seatbelt on macOS, bwrap + seccomp on Linux, MXC on Windows; .git, .codex, .agents stay read-only
MCP
CLI, IDE extension and ChatGPT desktop share one config; per-server tool allow-lists and approval modes
Extensions
Hooks ✓Subagents ✓Plugins ✓

MCP ↗AGENTS.md ↗Skills ↗Approvals & security ↗

Profile on AIMap →

Cursor (editor · CLI) Anysphere

Permissions
Run Modes: Auto-review (recommended, classifier) / Allowlist / Run Everything (no checks)
Sandbox
Shell commands can run sandboxed (Auto-review, or Allowlist with sandbox); CLI uses /sandbox or --sandbox. Cursor says Auto-review is not a security boundary
MCP
Or one-click install from the Cursor Marketplace in Customize; each MCP call needs approval unless allow-listed
Extensions
Hooks ✓Subagents ✓Plugins ✓

MCP ↗Rules ↗Skills ↗Run Modes ↗CLI ↗

Profile on AIMap →

GitHub Copilot (VS Code · Copilot CLI) GitHub

Permissions
CLI asks by default; --allow-tool grants specific tools, --allow-all-tools hands over all your access
Sandbox
/sandbox enable in the CLI for a local sandbox; cloud sandboxes are in public preview
MCP
Copilot CLI ships with GitHub’s MCP server; repo-level MCP config takes precedence over user config
Extensions
Hooks not documentedSubagents not documentedPlugins not documented

CLI MCP ↗VS Code MCP ↗Instructions ↗Skills ↗Copilot CLI ↗

Profile on AIMap →GitHub Copilot CLI →

Antigravity CLI Google

⚠ Launched 2026-05-19 as Gemini CLI’s successor; since 2026-06-18 Gemini CLI no longer serves free and Google AI Pro / Ultra individual users

Permissions
not documented
Sandbox
Has a dedicated sandbox / file-access page in the official docs
MCP
Remote servers must use serverUrl (url / httpUrl are not supported); /mcp opens the manager; install migrates Gemini CLI MCP servers and skills
Extensions
Hooks ✓Subagents ✓Plugins ✓

MCP ↗Skills ↗Migrate from Gemini CLI ↗Announcement ↗

Profile on AIMap →

Gemini CLI Google

⚠ Still served only for enterprise (Gemini Code Assist licences) and API-key users; individuals should move to Antigravity CLI

Permissions
not documented
Sandbox
Enable with -s / GEMINI_SANDBOX: docker, podman, macOS Seatbelt and more
MCP
trust: true skips every confirmation for that server—only for servers you fully control; sensitive env vars (*TOKEN*, *KEY*…) are stripped from MCP processes by default
Extensions
Hooks ✓Subagents ✓Plugins ✓

MCP ↗Transition notice ↗

Profile on AIMap →

Devin Desktop / Devin CLI (formerly Windsurf) Cognition

⚠ Windsurf became Devin Desktop on 2026-06-02; the default local agent moved from Cascade to Devin Local (shares Devin CLI config)

Permissions
not documented
Sandbox
Devin CLI has OS-level sandboxing (Linux needs bubblewrap + socat); enterprises can enforce it
MCP
The legacy Cascade agent keeps its own mcp_config.json
Extensions
Hooks ✓Subagents ✓Plugins ✓

Devin CLI MCP ↗Rules & AGENTS.md ↗Skills ↗Sandbox ↗Windsurf → Devin Desktop FAQ ↗

Profile on AIMap →

Cline Cline

Permissions
Asks step by step by default; Auto Approve can be enabled per action type
Sandbox
not documented
Extensions
Hooks not documentedSubagents ✓Plugins not documented

MCP ↗Rules ↗Skills ↗

Profile on AIMap →

Qwen Code Alibaba Qwen

Permissions
Approval modes plan / default / auto-edit / auto (classifier) / yolo
Sandbox
tools.sandbox can be set to docker and others
Extensions
Hooks not documentedSubagents not documentedPlugins not documented

MCP ↗Skills ↗Settings ↗

Profile on AIMap →

04FAQ

Which rules file works in the most agents?

AGENTS.md: 9 of the 11 clients here read it (Claude Code, Codex, Cursor, GitHub Copilot, Antigravity CLI, Devin Desktop / Devin CLI, Cline, OpenCode, Kiro). Put shared rules there and keep tool-specific files thin.

Where should I put Agent Skills so several agents find them?

.agents/skills/<name>/SKILL.md — 7 of 11 clients scan it. Claude Code reads .claude/skills/, which Cursor, Copilot, Cline and OpenCode also read for compatibility.

Can I copy an MCP JSON block from one client to another?

Mostly, with fixes: Claude Code needs "type": "http" on any entry with a url; Antigravity CLI uses serverUrl instead of url/httpUrl; Codex uses TOML ([mcp_servers.<name>]); OpenCode uses a "mcp" key with type local/remote.

Is Gemini CLI still usable?

Since 2026-06-18 Google no longer serves free and Google AI Pro / Ultra individual users in Gemini CLI; enterprise licences and API-key users still work. Individuals are pointed to Antigravity CLI, which imports skills, MCP servers and GEMINI.md.

About this data · for AI to read

Data updated: Agent setup guide(UTC+8)