The Grammar of Coding Agents

OpenAI Codex, Claude Code, OpenCode, Cursor, and GitHub Copilot

Snapshot verified: June 19, 2026
Scope: Interactive command languages, context sigils, skills, subagents, planning, review, insights, and differences between terminal, desktop, IDE, and cloud harnesses.

This is a living reference. Command availability can vary by version, operating system, plan, experimental flags, and connected extensions. In every harness, typing / in the active composer is the best final check for what is available in that exact installation.

Executive summary

The five product families increasingly share the same building blocks - slash commands, reusable skills, project instructions, MCP, custom agents, shell execution, and context management - but they expose them through different command languages.

• Codex has the clearest use of $: $skill-name explicitly invokes a skill, while $app-slug attaches an app/connector. @ is primarily workspace/file context, though recent Codex releases are moving toward a unified mentions menu. Codex has /goal, but parallel subagents are requested in natural language rather than through /fleet.
• Claude Code has the broadest built-in command catalog and the richest collection of packaged workflows. Skills are invoked as /skill-name, not with $. Its stand-out commands include /insights, /batch, /code-review, /loop, /ultraplan, and /workflows.
• OpenCode has a smaller built-in slash-command surface but a very flexible customization model. @ can reference both files and agents; custom commands support $ARGUMENTS, $1, $2, shell-output injection, file references, agent selection, and forced subtask execution.
• Cursor combines an IDE, terminal CLI, and Cloud Agents around one rules, skills, MCP, and subagent system. Skills and custom subagents use /name; @ attaches files, folders, docs, terminals, prior chats, diffs, browser state, and manual rules. Cursor CLI uses /shell rather than a documented ! shortcut, and & prompt hands a local CLI task to a Cloud Agent.
• GitHub Copilot must be split by harness. Copilot CLI and the Copilot app use a modern agent runtime with /fleet, /chronicle, /research, skills, plugins, and custom agents. VS Code Chat uses a different interaction model centered on /explain, /fix, #file, and @workspace. GitHub.com Chat has another, smaller set of conversation commands and @ attachments.

How to read the comparison

The closest equivalents are:

1. The sigils: what /, @, $, #, and ! mean

Sigil portability at a glance

Practical consequence

The most dangerous portability mistake is assuming the same sigil means the same thing:

• @workspace is a Copilot VS Code participant, not a portable way to request repository context elsewhere.
• $frontend-design is a Codex skill invocation, but Claude Code and Copilot normally expect /frontend-design.
• $ARGUMENTS in Claude Code and OpenCode belongs inside the reusable command definition; it is not how an end user invokes the command.
• #file is Copilot vocabulary. Codex, Claude Code, and OpenCode normally use @file.
• Cursor uses /skill-name and /subagent-name; its @ menu is for context and manual rules, not the primary custom-agent selector.

2. Capability comparison

Legend: Yes = directly documented; Partial = available through another mechanism or only some surfaces; No = no close built-in equivalent documented.

3.OpenAI Codex

Harnesses

High-value Codex CLI commands

Codex app slash commands

/feedback, /goal, /init, /mcp, /plan, /review, /status

Codex IDE extension slash commands

/auto-context, /cloud, /cloud-environment, /feedback, /goal, /local, /review, /status

Skills, apps, and context

• Use $skill-name to explicitly invoke a skill.
• Use /skills to browse available skills; enabled skills also appear in the slash menu.
• /apps inserts a connector as $app-slug.
• Use @path or /mention for file context.
• Use !command for a local shell command.

Subagent behavior

Codex supports parallel subagents in the CLI and app, but the trigger is an explicit natural-language request rather than /fleet. For example:

Spawn three subagents in parallel: one for security, one for test gaps,
and one for maintainability. Wait for all three and merge the findings.

/agent does not itself mean “start a fleet”; it switches between already spawned agent threads.

Distinctive strengths

• Persistent /goal state with pause/resume/clear behavior.
• First-class $ syntax for skills and apps.
• Tight local/cloud handoff across CLI, app, and IDE.
• App-centric parallel threads, worktrees, automations, artifact previews, and Git review.

4.Claude Code

Harnesses

Claude Code is available in the terminal, IDE, desktop app, and browser. Anthropic describes the CLI as the most complete surface; desktop and IDE trade some CLI-only capabilities for visual review and tighter editor integration.

High-value Claude Code commands

Commands that are especially distinctive

• /insights analyzes Claude Code sessions for project areas, interaction patterns, and friction points.
• /batch decomposes a large change into roughly 5-30 worktree-isolated subagents, each expected to implement, test, and open a pull request.
• /code-review can review, fix, comment on a PR, or run an ultra cloud review.
• /simplify runs multiple cleanup-oriented review agents.
• /loop runs a repeated or self-paced task.
• /ultraplan and /ultrareview connect terminal work to deeper cloud planning/review.
• /team-onboarding analyzes recent usage and generates a teammate onboarding guide.

Skills and custom commands

• Invoke a skill as /skill-name.
• User skills live under .claude/skills/<name>/SKILL.md.
• Older .claude/commands/*.md files still work, but Anthropic now treats custom commands as skills.
• $ARGUMENTS and positional parameters are used inside skill definitions.
• MCP prompts appear as /mcp__<server>__<prompt>.

Context and shell syntax

• @path opens file-path autocomplete and attaches file context.
• !command runs a command directly and adds its output to the conversation.
• In the multi-agent view, @agent-name can choose a custom agent and @repo can target a sibling repository.

Subagent choices

Claude Code has several distinct parallelism mechanisms:

1. Subagents for isolated delegated tasks.
2. Agent view (claude agents) for multiple background sessions.
3. Agent teams for coordinated sessions that communicate.
4. Dynamic workflows for background multi-phase orchestration.
5. /batch as a packaged large-change workflow using worktrees and subagents.

Distinctive strengths

• The broadest command vocabulary.
• Strong history analysis through /insights.
• Several levels of parallel execution, from one-off subagents to teams and workflows.
• Rich packaged workflows for review, debugging, research, verification, onboarding, and scheduling.

5.OpenCode

Harnesses

OpenCode is available as a TUI, desktop app, IDE extension, and browser application. It also supports ACP clients. ACP carries most OpenCode capabilities, although some built-in commands such as /undo and /redo are not supported there.

Built-in TUI commands

Custom commands

OpenCode's small built-in list is offset by a strong custom-command format:

• Define Markdown commands in .opencode/commands/.
• Use $ARGUMENTS, $1, $2, and so on for parameters.
• Inject shell output into a command prompt.
• Include files with @path.
• Pin a command to an agent or model.
• Set subtask: true to force the command into a separate subagent context.
• A custom command can override a built-in command with the same name.

Example:

---
description: Review one component
agent: plan
subtask: true
---

Review @$1 for correctness, accessibility, and performance.
Run the smallest relevant tests and summarize the risks.

Agents

• Primary agents: Build and Plan.
• Switch primary agents with Tab or the configured agent-switch key.
• Built-in subagents: General, Explore, and Scout.
• Invoke a subagent directly with @general, @explore, or @scout.
• Primary agents may also delegate automatically based on subagent descriptions.

Skills

OpenCode supports SKILL.md skills and searches .opencode/skills, .claude/skills, and .agents/skills at project and user scope. Skills are discovered by metadata and loaded on demand through the native skill tool.

Context and shell syntax

• @path attaches a file or configured reference.
• @agent can invoke a subagent.
• !command runs a shell command and adds the result to context.
• $... is primarily template syntax inside custom commands, not a first-class end-user skill selector.

Distinctive strengths

• Provider-agnostic and open source.
• Very clean Build/Plan split.
• Custom commands can select model, agent, and subtask behavior.
• Compatibility with Claude-style and Agent Skills directories makes it easy to share reusable behavior.

6.Cursor

Cursor is best understood as a family of connected harnesses rather than only an editor. The editor, terminal CLI, Agents Window, Cloud Agents, and Bugbot share rules, skills, MCP configuration, custom subagents, and session context to different degrees.

Harnesses

High-value Cursor CLI commands

Cursor CLI also supports:

• agent -p or agent --print for non-interactive output.
• --force or --yolo when a headless run is allowed to modify files.
• --worktree [name] or -w for an isolated worktree.
• agent ls, agent resume, --continue, and --resume=<id> for session recovery.
• agent acp when Cursor is used through an ACP-compatible host.

Context, plans, and worktrees

• Use @file or @folder to attach repository context.
• The editor also exposes @Docs, @Terminals, @Past Chats, Git working-state or branch diffs, and @Browser.
• Plan mode is available from Shift+Tab, /plan, --plan, or --mode=plan.
• Ask mode is available as /ask or --mode=ask.
• The editor supports /worktree, /best-of-n, /apply-worktree, and /delete-worktree; the CLI uses --worktree.
• In the CLI, Ctrl+R opens a review of the agent's changes.

Skills and rules

• Skills use SKILL.md and can be selected automatically or explicitly as /skill-name.
• disable-model-invocation: true makes a skill explicit-only.
• Cursor discovers skills from .agents/skills, .cursor/skills, .claude/skills, and .codex/skills, including supported user-level locations.
• /migrate-to-skills converts older command-style customizations.
• Project rules live in .cursor/rules/*.mdc; Cursor also supports AGENTS.md, user rules, and team rules.
• A manually applied rule can be attached with @rule-name; /create-rule helps create a rule.

Subagents

• Cursor can delegate automatically based on a subagent's description.
• Explicit invocation uses /subagent-name or a natural-language request.
• Multiple independent subagent tasks can run in parallel.
• Built-in subagents include Explore, Bash, and Browser.
• Custom agents live in .cursor/agents; Cursor also reads compatible .claude/agents and .codex/agents locations.
• Agent definitions can specify a model, read-only behavior, and background execution.
• Background subagent output is stored under ~/.cursor/subagents/.

Cloud handoff and review

• Prefix a Cursor CLI prompt with & to send it to a Cloud Agent.
• /in-cloud sends the next delegated task to a cloud subagent.
• /babysit monitors and helps shepherd a pull request.
• /review and /review-bugbot are documented for Cursor Agent and cursor.com/agents in Cursor 3.7 and later. Cursor's documentation currently says CLI support is coming later.
• Bugbot Autofix can create a Cloud Agent task from a review finding.

Distinctive strengths

• One customization model spans editor, CLI, local agents, and cloud agents.
• The richest @ context vocabulary in this comparison.
• Explicit /skill-name and /subagent-name syntax keeps reusable behavior easy to discover.
• Native local worktrees, best-of-N execution, cloud handoff, PR review, and browser-aware work.

7.GitHub Copilot: compare the harnesses separately

Copilot's surfaces are not one command language

7.1GitHub Copilot CLI

Copilot CLI now has a large terminal-agent command surface.

/fleet

/fleet decomposes work and runs independent subtasks in parallel. Copilot can choose custom agents for subtasks, and you can name one in the prompt with @custom-agent-name.

/fleet Implement the API validation, add integration tests, and update the docs.
Use @test-writer for the test work.

/chronicle

/chronicle is Copilot CLI's closest equivalent to Claude Code's /insights.

Documented subcommands include:

• standup - summarize recent work.
• tips - personalized workflow suggestions.
• cost tips - token-spend feedback.
• search - search session history.
• improve - suggest improvements to custom instructions.
• reindex - rebuild and synchronize the session store.

The exact short form shown by /help may vary as the feature evolves.

Skills

• Copilot may automatically select skills.
• Explicitly request one with /skill-name.
• Manage them with /skills list, /skills info, /skills add, /skills remove, and /skills reload.

7.2GitHub Copilot app

The desktop app is built on the Copilot CLI runtime and is designed for managing multiple local and cloud agent sessions. It shares session-history data with the CLI and can contribute to /chronicle reports.

Important distinction: runtime sharing does not guarantee that every CLI slash command is exposed identically in every app composer. Use the app's live / menu as the authoritative list.

Best fit:

• Multiple concurrent agent sessions.
• Visual session management.
• GitHub-native repository, branch, CI, issue, and PR workflows.
• MCP servers configured from repositories, CLI configuration, or app settings.

7.3Copilot Chat on GitHub.com

GitHub.com Chat uses a lighter command language:

• /clear, /delete, /new, /rename.
• @ attaches discussions, extensions, files, issues, pull requests, and repositories.
• GitHub MCP skills can perform repository operations from natural language.

This is a GitHub-object chat surface, not a local terminal harness.

7.4Copilot Chat in VS Code

Common slash commands:

/clear, /explain, /fix, /fixTestFailure, /help, /new, /tests

Common chat variables:

#block, #class, #comment, #file, #function, #line, #path, #project, #selection, #sym

Common chat participants:

@azure, @github, @terminal, @vscode, @workspace

Useful query patterns:

@workspace Explain the authentication flow and identify the trust boundaries.

#file:src/auth.ts /tests Generate tests for the token refresh edge cases.

@github #web What is the current supported Node.js LTS, and what changes would
this repository need to upgrade?

Prompt files under .github/prompts/*.prompt.md create reusable /prompt-name commands in supported IDEs.

7.5Copilot cloud agent

Copilot cloud agent runs independently in a GitHub-hosted environment, works on a branch, and can create a pull request. It can be started from GitHub, mobile, IDEs, API, GitHub CLI, the GitHub MCP Server, and external work-management integrations.

Do not confuse:

• IDE agent mode: acts in the local workspace.
• Copilot cloud agent: acts in a managed GitHub environment.
• Copilot CLI: acts locally unless explicitly delegated or started with a cloud mode.

8. Query translation guide

9. Recommendations by working style

10. Other coding agents and harnesses to add later

This is a candidate backlog, not an exhaustive market inventory.

Terminal-first or terminal-capable

• Google Antigravity CLI - Google's consumer successor to Gemini CLI, sharing an agent harness with Antigravity 2.0, IDE, and SDK surfaces
• Aider
• Amp
• Cline CLI
• Continue CLI
• Crush
• Plandex
• Pi Coding Agent
• Goose
• Junie CLI
• Kiro CLI
• Factory Droid

IDE/editor-first

• Devin Desktop / Cascade (the product formerly known as Windsurf)
• Cline
• Roo Code
• Continue
• JetBrains Junie
• Kiro IDE
• Zed Agent Panel
• Augment Code
• Replit Agent

Cloud/issue-to-PR and engineering-agent platforms

• Devin
• OpenHands
• Factory
• GitLab Duo Agent Platform
• Replit Agent
• Augment Remote Agents
• Warp cloud agents / agent control plane

Useful adjacent harness hosts

• Warp - can run its own agent and host external CLI agents.
• Zed with ACP - editor host for external agents.
• VS Code Agent mode - a harness layer that can host Copilot and third-party agents.

Recommended next comparison rows

For the next revision, the most useful additions would be:

1. Google Antigravity - compare its CLI, IDE, Antigravity 2.0 desktop command center, SDK, shared multi-agent harness, skills, hooks, plugins, and asynchronous subagents.
2. Cline and Roo Code - model-agnostic VS Code agents with modes, rules, MCP, and skills.
3. Aider - mature terminal command language and Git-centered workflow.
4. Kiro and Junie - spec-driven and JetBrains-native agent approaches.
5. Devin, Factory, and OpenHands - cloud/autonomous engineering-agent category.

11. Maintenance protocol

When updating this report:

1. Record the verification date at the top.
2. Re-check each product's official command reference and live / menu.
3. Separate stable, preview, and experimental commands.
4. Keep GitHub Copilot surfaces in separate rows.
5. Keep Cursor editor, CLI, Agents Window, Cloud Agents, and Bugbot distinctions explicit.
6. Verify whether a command is a built-in command, a bundled skill, a custom skill, an MCP prompt, or a UI action.
7. Re-check the meanings of /, @, $, #, and !.
8. Update the “other agents” backlog for renames, retirements, and replacements.
9. Add a short change-log entry below.

Change log

Sources

All core comparisons use official documentation current on June 19, 2026.

OpenAI Codex

1. Codex CLI slash commands
2. Codex app commands
3. Codex IDE slash commands
4. Codex Agent Skills
5. Codex subagents
6. Codex CLI features
7. Codex app features

Claude Code

8. Claude Code commands
9. Claude Code skills
10. Claude Code subagents
11. Claude Code parallel agents
12. Claude Code interactive mode
13. Claude Code platforms

OpenCode

14. OpenCode TUI
15. OpenCode commands
16. OpenCode agents
17. OpenCode Agent Skills
18. OpenCode ACP support

Cursor

19. Cursor Agent overview
20. Cursor prompting and @ context
21. Cursor Plan mode
22. Cursor CLI overview
23. Using Cursor CLI
24. Cursor CLI slash-command reference
25. Cursor Agent Skills
26. Cursor subagents
27. Cursor Cloud Agents
28. Cursor rules
29. Cursor Bugbot
30. Cursor worktrees

GitHub Copilot

31. Copilot CLI command reference
32. Copilot Chat cheat sheet
33. Copilot CLI fleet
34. Copilot CLI session data and Chronicle
35. Copilot CLI Agent Skills
36. GitHub Copilot app
37. Copilot cloud agent
38. Copilot customization cheat sheet

Candidate-agent official pages

39. Google Antigravity documentation
40. Google's Gemini CLI to Antigravity transition announcement
41. Aider
42. Cline
43. Roo Code
44. Continue
45. Goose
46. Junie
47. Kiro
48. Devin
49. Factory
50. OpenHands
51. Replit Agent
52. Amp
53. Zed Agent Panel
54. Warp
55. Augment Agent
56. Crush
57. Plandex
58. Pi Coding Agent
59. GitLab Duo Agent Platform