.agents/ for all your tools
We all want to give our AIs the right context at the right time. But while there are standards, there are as many different approaches as there are tools.
We believe that developers should only have to define their AI Workflow once, and have it work across all tools.
Support matrix
How well each tool supports MCP configuration, AGENTS.md, and agent skills from a unified project location. Click any row for details.
| Tool | AGENTS.md | MCP | Skills | Notes | Issues |
|---|---|---|---|---|---|
 Claude Code▾ | Best-in-class deterministic progressive disclosure, but only for CLAUDE.md. AGENTS.md not natively supported. --add-dir adds skill locations with live reloading but requires invocation each start. | — | |||
 Cursor▾ | — | ||||
 Devin Desktop (Windsurf)▾ | Recommended skill path is .agents/skills/<skill-name>/SKILL.md. Org-level skills repo is the intended workaround for global skills. | — | |||
 Gemini CLI▾ | AGENTS.md must be configured as an alternative filename in settings.json to be picked up. | — | |||
 GitHub Copilot▾ | — | ||||
 Kiro▾ | Kiro Steering inclusion modes control AGENTS.md. Skills require manual one-time import; subagents do not automatically inherit skills. | — | |||
 OpenCode▾ | AGENTS.md support limited to project root and directories up to repo root only. | — | |||
 Visual Studio Code▾ | Allows configuring skill locations. AGENTS.md progressive disclosure only by model request. | — |
Standardize today
Until tools adopt the spec natively, these lightweight npm packages let you maintain context in .agents/and sync it automatically to where each tool expects it. You can automate with scripts or hooks to provide all team members and CI/CD tools with the context they need, and seamlessly hover between the coding tools of your choice.
Reads .agents/mcp.json and copies MCP server config to the paths each AI coding tool expects natively.
Ensures every AGENTS.md file in your project has a corresponding vendor context file (e.g. CLAUDE.md) that references it.
Copies agent skills from .agents/skills/ to the vendor-specific directories each tool reads from.
No matter the vendor, it's agents.
Dotagents proposes a single, portable location for AI coding context: the .agents/ directory at your project root. This is not a new idea; many tools already support it.
Place your AGENTS.md files, MCP server configuration, and agent skills there once. Every AI coding tool that adopts the spec will pick them up automatically, regardless of vendor.
Directory structure
project-root/
├── AGENTS.md ← project-wide rules
├── .agents/
│ ├── AGENTS.md ← same, alternative location
│ ├── mcp.json ← MCP server definitions
│ └── skills/
│ ├── my-skill/
│ │ └── skill.md
│ └── other-skill/
│ └── skill.md
└── src/
└── feature/
└── AGENTS.md ← scoped to this subtreeProgressive disclosure
Subdirectory AGENTS.md files are more powerful when the harness includes them automatically based on which files are touched in a session. This is deterministic and reliable. When the model has to request additional context files itself, it may do that. It also may not.
Context file weight
Models tend to follow their own vendor context files (CLAUDE.md,GEMINI.md) more reliably than generic files, even when AGENTS.md is included in the chat. These vendor-agnostic context files should be picked up with the same weight as vendor-specific ones.
Skills & MCP Servers
Skills and MCP Servers have special lifecycles and meanings across tools. While they should be natively supported and live-reloaded from the .agents/ directory, progressive disclosure is not always reasonable and should only be expected when the tool supports it for its own, native implementations too. MCP Servers should generally disclose by model request much like skills do.
More ideas? The spec is lightweight and open to contribution.
Read the spec on GitHub →Do you want this too?
Here is how you can help, for free. Every action, however small, contributes to getting AI coding tools to adopt a shared standard.
Find the open issue for your preferred tool and vote or comment to increase visibility.
Browse issues →Write about it, post on Reddit, or mention it in discussions where developers talk about AI coding tools.
Missing a tool? Found an error in the support matrix? Open a pull request or issue.
Open an issue →Help improve xmcp, xagents, or xskills — small TypeScript packages with focused scope.
xmcp →This is a solo, non-profit community initiative. Ko-fi support helps cover hosting and domains.
Support on Ko-fi →