Claude joining the list of tools that read AGENTS.md sounds like unification. It is really the opposite: now every AI coding assistant claims to honor the same file, but each interprets it slightly differently. Cursor still reads .cursorrules, Aider reads .aider.conf.yml, JetBrains has its own layer, and enterprise tools quietly append policies to whatever they find. We tried seven of the most-used AI coding apps to see how they treat AGENTS.md, where they diverge, and which ones give you a clean way to keep your rules in one place.
What to look for in an AGENTS.md-aware app
The point of a shared convention is fewer files. The tools worth adopting share these traits.
- Reads the top-level
AGENTS.mdwithout extra configuration - Honors nested
AGENTS.mdinside a subdirectory when the agent is working there - Documents which sections the model actually consumes (some tools use only the first heading)
- Exposes a preview of what the tool sent to the model
- Works with monorepos where per-package rules matter more than a global file
- Does not silently rewrite the file on its own
Quick comparison
| App | Best for | Free plan | Standout feature |
|---|---|---|---|
| Aider | Terminal-first coding with git diffs | Yes | Reads AGENTS.md and .aider.conf.yml together |
| Claude Code | Long-running tasks with an agentic loop | Free tier | Honors AGENTS.md, plugin skills, and hooks |
| Codex CLI | OpenAI-native terminal workflow | Yes | Uses AGENTS.md as the primary policy |
| Cursor | IDE with tab-completion and chat | Free tier | Bridges .cursorrules and AGENTS.md |
| Continue | Editor extension with a local backend | Yes | Reads AGENTS.md through its context providers |
| Zed | Fast editor with first-class AI panes | Yes | AGENTS.md loads as a system prompt |
| Windsurf | Managed IDE from the Codeium team | Free tier | Custom rules panel plus AGENTS.md |
The apps
1. Aider — best terminal-first pick
Aider is the most literal reader of AGENTS.md. Point it at a repo and it treats the file as the system prompt, layering .aider.conf.yml on top for tool-specific behavior. Git integration is native: every edit lands as a commit you can review or revert.
Where it falls short: No IDE integration by default. Configuration lives in dotfiles rather than a panel.
Platforms: Python on Linux, macOS, Windows.
Pricing: Free, Apache 2.0.
Download: aider.chat
Bottom line: The right pick when the terminal is the IDE and Git is the review.
2. Claude Code — best for long-running work
Claude Code reads AGENTS.md alongside its own CLAUDE.md, its plugin skills, and repo hooks. The agent loop keeps the file in context across many tool calls, which matters when the rules include long lists of do-nots. Nested AGENTS.md files in a subdirectory are honored when the agent’s cwd moves.
Where it falls short: Nested-file precedence when both AGENTS.md and CLAUDE.md exist takes a minute to grok. Some plugins add hidden policy that is not obvious from the repo alone.
Platforms: Linux, macOS, Windows terminal. Also available in the browser via claude.ai/code.
Pricing: Free tier available. Paid tier for higher usage.
Download: claude.com/code
Bottom line: The right pick for multi-hour agentic runs where the rulebook has to stay applied.
3. Codex CLI — best for OpenAI-native workflows
Codex CLI was one of the first shells to standardize on AGENTS.md, and it still treats the file as the primary policy source. There is no second rules layer to keep in sync, which is either refreshing or limiting depending on your setup.
Where it falls short: Fewer knobs than Aider or Claude Code. Extensibility depends on OpenAI’s release cadence.
Platforms: Node CLI on Linux, macOS, Windows.
Pricing: Free CLI, uses your OpenAI API credits.
Download: github.com/openai/codex
Bottom line: The right pick if OpenAI is your model and you want minimal rule-file surface.
4. Cursor — best IDE that reads AGENTS.md
Cursor used to read .cursorrules and only that. In 2026 it also reads AGENTS.md, layered on top of the older rules file. Nested rules inside a workspace folder work well, and the “rules” panel now shows exactly what the model received.
Where it falls short: Two rules formats coexisting invites drift. Some features still key off .cursorrules even when AGENTS.md says otherwise.
Platforms: Windows, macOS, Linux. VS Code-based.
Pricing: Free tier. Pro at $20 monthly.
Download: cursor.com
Bottom line: The right pick when a graphical IDE is non-negotiable and you already use Cursor.
5. Continue — best VS Code extension
Continue is a VS Code and JetBrains extension that lets you point at Ollama, LocalAI, or a cloud provider. Its context providers include an AGENTS.md reader, and its config file lets you compose the file with per-command overrides. The extension shows the composed prompt before sending, which is what most developers actually want.
Where it falls short: The composed-prompt preview is powerful but takes time to learn. Local models still trail cloud ones for hard refactors.
Platforms: VS Code and JetBrains on Linux, macOS, Windows.
Pricing: Free, Apache 2.0.
Download: continue.dev
Bottom line: The right extension when you want editor-native AI plus a local backend.
6. Zed — best fast editor with AI panes
Zed loads AGENTS.md as the system prompt for its assistant panels. The rest of the editor is fast in a way most VS Code forks are not. If the codebase is large and the review loop is tight, Zed’s responsiveness compounds.
Where it falls short: Windows support arrived later than Linux and macOS. Fewer AI features than dedicated IDEs.
Platforms: macOS, Linux, Windows.
Pricing: Free tier. Optional paid plans for hosted model access.
Download: zed.dev
Bottom line: The right pick when editor speed matters more than the biggest AI feature list.
7. Windsurf — best managed IDE
Windsurf is the IDE from the Codeium team. It has a dedicated rules panel that layers cleanly with AGENTS.md, and its agentic runs (Cascade) keep the file in context across multi-file changes. Enterprise policies can be pushed centrally without editing the repo file.
Where it falls short: Not open source. The paid tier is the value tier; the free plan is enough to try it.
Platforms: Windows, macOS, Linux.
Pricing: Free tier. Pro plans for individuals and teams.
Download: windsurf.com
Bottom line: The right pick when a team wants central policy on top of AGENTS.md.
How to pick the right one
- If the terminal is where you code: Aider or Codex CLI
- If the workload is multi-hour agentic tasks: Claude Code
- If Cursor is already the daily driver: keep it and let it read AGENTS.md alongside
.cursorrules - If the goal is a local backend inside VS Code: Continue
- If editor speed matters more than AI feature depth: Zed
- If a team needs central rule policy: Windsurf
FAQ
Should I write CLAUDE.md, AGENTS.md, or both?
Write AGENTS.md first. If you use Claude Code specifically, add a small CLAUDE.md for Claude-only overrides. Nothing in the ecosystem punishes you for having both.
Does AGENTS.md work in monorepos?
Yes. Every tool on this list respects a nested AGENTS.md, though the precedence rules differ. Aider and Claude Code prefer the deepest one; Cursor merges upward from the root.
How do I stop a tool from silently editing my rules?
Set the file to read-only in the repo or, for extra safety, generate it from a template kept in docs/. Some tools honor a “do not edit” comment at the top; most now do.
Which tool preview shows me what the model actually receives? Continue and Cursor both show the composed prompt. Claude Code prints the tool call trail when verbosity is turned up.
Does AGENTS.md work with local models? Yes, as long as the tool honors it. Continue plus Ollama, and Aider plus a local model both apply the file correctly.
Is there a linter for AGENTS.md? Nothing widely adopted yet. Teams typically enforce structure with a simple pre-commit hook that checks for required sections.
