Agent Portability

AllAgents syncs agent files (.claude/agents/ to .github/agents/, etc.) verbatim — no transformation is needed. Clients ignore frontmatter fields they don’t recognize, so the same file works everywhere.

However, some frontmatter fields are client-specific and won’t be interpreted by other clients. This guide covers how to write agents that work well across all clients.

How Agent Sync Works

When a plugin contains an agents/ directory, AllAgents copies those files to each client’s agent path:

Client	Agent Path
Claude	.claude/agents/
Copilot	.github/agents/

The files are copied (or symlinked) verbatim. No frontmatter transformation occurs because clients safely ignore fields they don’t understand.

Writing Portable Agents

Universal Fields

These frontmatter fields are understood by both Claude Code and GitHub Copilot:

Field	Description
name	Agent identifier
description	What the agent does and when to use it

Shared Fields (Different Syntax)

These fields exist in both clients but use client-specific values:

Field	Claude Code	Copilot
tools	Read, Grep, Glob, Bash, Edit, Write	read, edit, search/codebase, web/fetch
model	sonnet, opus, haiku, inherit	Claude Opus 4.5, GPT-5 (copilot)
MCP servers	mcpServers (camelCase)	mcp-servers (kebab-case)
hooks	Supported	Supported

Client-Specific Fields

These fields are valid in their respective clients but ignored by others:

Claude Code only:

• disallowedTools, permissionMode, maxTurns, skills, memory, background, effort, isolation, initialPrompt

GitHub Copilot only:

• disable-model-invocation, user-invocable, target, metadata, agents, argument-hint, handoffs

Best Practices

For maximum portability:

1. Keep name and description — these are universal
2. Omit tools — each client will use its full default toolset, which is usually what you want
3. Omit model — each client will use its default model
4. Put behavior in the markdown body — the system prompt (everything after the frontmatter) is the strongest interop point; all clients read it identically

---
name: code-reviewer
description: Reviews code for quality, security, and best practices. Use proactively after code changes.
---

You are a senior code reviewer. When invoked:

1. Check recent changes with git diff
2. Review modified files for quality, security, and maintainability
3. Provide feedback organized by priority (critical, warnings, suggestions)

This agent works identically in Claude Code and Copilot — both understand the name and description, and both follow the system prompt instructions using whatever tools they have available.

When Client-Specific Fields Are Fine

If you only target one client, use any fields you need. Claude-specific fields like tools: Read, Grep, Glob or model: haiku are harmless in Copilot — they’re simply ignored. The agent still works; it just uses Copilot’s defaults instead.

---
name: safe-researcher
description: Read-only research agent for codebase exploration
tools: Read, Grep, Glob, Bash
model: haiku
---

You are a research agent. Explore the codebase and report findings.
Do not modify any files.

In Claude Code, this agent is restricted to read-only tools and uses Haiku. In Copilot, the tools and model fields are ignored, so the agent has full tool access and uses Copilot’s default model. The system prompt instruction “Do not modify any files” still guides behavior in both.

Multi-Provider Resilience

Running multiple AI coding tools in parallel provides resilience against model provider outages. If one provider goes down, your team can switch to another tool that’s already configured with the same agents, skills, and project context.

AllAgents makes this seamless by syncing your configuration across all clients simultaneously.

Setting Up Dual-Provider Workspaces

Configure your workspace with multiple clients:

.allagents/workspace.yaml

plugins:
  - code-review@my-org/ai-plugins
  - my-org/shared-agents

clients:
  - claude
  - copilot

Running allagents update syncs agents, skills, hooks, and MCP servers to both clients. Your team can use either tool interchangeably.

What Gets Synced

Artifact	Claude	Copilot
Skills	.claude/skills/	.github/skills/
Agents	.claude/agents/	.github/agents/
Hooks	.claude/hooks/	.github/hooks/
Commands / Prompts	.claude/commands/	.github/prompts/
Agent file	CLAUDE.md	AGENTS.md
MCP servers	.claude/mcp.json	.copilot/mcp-config.json
WORKSPACE-RULES	Injected in CLAUDE.md	Injected in AGENTS.md

Skills Are the Strongest Interop Point

Skills (reusable prompt files in skills/ directories) are read natively by both Claude Code and Copilot. They use the same format and the same frontmatter. When writing shared plugins, prefer skills over client-specific commands or hooks.

Skill Frontmatter Portability

Unlike agents, skills have strong frontmatter alignment across clients:

Field	Claude Code	Copilot	Portable
name	Required	Required	Yes
description	Required	Required	Yes
disable-model-invocation	Supported	Supported	Yes
user-invocable	Supported	Supported	Yes
argument-hint	Not supported	Supported	No (Copilot only)
allowed-tools	Supported	Not supported	No (Claude only)
model	Supported	Not supported	No (Claude only)

For portable skills, use name, description, disable-model-invocation, and user-invocable. Avoid allowed-tools and model — these are Claude-specific and ignored by Copilot.

Resilience Without AllAgents

Without AllAgents, teams maintain separate configurations for each tool — duplicate AGENTS.md and CLAUDE.md files, manual sync-agents.sh scripts, and per-tool MCP registration. This duplication is error-prone and drifts over time.

AllAgents eliminates this by treating your workspace.yaml as the single source of truth and syncing to all clients on every allagents update.