Overview
Cross-platform comparison
Cross-Platform Overview
Prompt-first procedure: Describe the outcome you want in your agent conversation. The agent should select and load the appropriate AIWG assets, explain material changes, request any needed approval, and report verification evidence. Exact commands and flags appear only in the CLI reference.
AIWG gives your AI assistant reusable project context and specialist workflows inside the AI tools you already use. The provider integration decides where that context, skills, commands, rules, and agents are written so the selected tool can read them. AIWG has 18 named provider integrations plus a separate provider-neutral `generic` fallback adapter.
Quick Comparison
The technical registry currently covers 18 named provider integrations, including Google Antigravity CLI, Grok Bot, Grok Build, Muse Code, Oh My Pi, and Pi Coding Agent, plus a provider-neutral `generic` fallback adapter. Start with Install, Connect, and Verify, then use the row below for provider-specific deployment details. The commands in this table are focused SDLC examples for manual use; the canonical installation path uses the complete supported setup unless project policy selects a narrower scope.
| Platform | Deploy Command | Context File |
|---|---|---|
| Google Antigravity CLI | the agent-owned use operation (alias `agy`) | AGENTS.md |
| Claude Code | the agent-owned use operation | CLAUDE.md |
| OpenAI/Codex | the agent-owned use operation | AGENTS.md |
| GitHub Copilot | the agent-owned use operation | copilot-instructions.md |
| Cursor | the agent-owned use operation | .cursor/rules/ (MDC) |
| DeepSeek Harness | the agent-owned use operation | AGENTS.md |
| Factory AI | the agent-owned use operation | AGENTS.md |
| Grok Bot | the agent-owned use operation | AGENTS.md |
| Grok Build | the agent-owned use operation | AGENTS.md |
| Muse Code | the agent-owned use operation | AGENTS.md |
| Hermes | the agent-owned use operation | AGENTS.md |
| OpenCode | the agent-owned use operation | AGENTS.md |
| OpenClaw | the agent-owned use operation | AGENTS.md |
| OpenHuman | the agent-owned use operation | Provider-managed |
| Oh My Pi | the agent-owned use operation | .omp/AGENTS.md |
| Pi Coding Agent (pi.dev) | the agent-owned use operation | AGENTS.md |
| Warp Terminal | the agent-owned use operation | WARP.md |
| Devin Desktop | the agent-owned use operation | AGENTS.md |
| Custom or unsupported harness | the agent-owned use operation | AGENTS.md |
What Gets Deployed
Artifacts deploy in each provider's native or compatibility format when that provider exposes the corresponding surface:
- Agents - Specialized AI personas (Architecture Designer, Test Engineer, Security Auditor, etc.)
- Commands - Slash commands and CLI commands (`/mention-wire`, `transition`, `where-are-we`)
- Skills - Natural language workflows (project awareness, handoffs, quality gates)
- Rules - Context rules and coding standards (citation policy, token security, versioning)
- Behaviors - Platform behavior definitions (OpenClaw only)
Provider Capability Matrix
| Provider | Agents | Commands | Skills | Rules | Behaviors |
|---|---|---|---|---|---|
| Google Antigravity CLI | degraded | indexed | native | AGENTS.md | unsupported |
| Claude Code | native | native | native | native | - |
| OpenAI/Codex | native | native | native | conventional | - |
| GitHub Copilot | native | native | conventional | native | - |
| Cursor | conventional | conventional | native | native | - |
| Factory AI | native | native | native | conventional | - |
| Hermes | aggregated | indexed | native | AGENTS.md + CLI | indexed |
| OpenCode | native | native | conventional | conventional | - |
| OpenClaw | native | native | native | native | native |
| OpenHuman | indexed | indexed | native | indexed | indexed |
| Oh My Pi | native | native prompts | native | native | extension bridge |
| Pi Coding Agent | skills-as-agents | native prompts | native | AGENTS.md | extension bridge |
| Warp Terminal | aggregated | conventional | native | aggregated | - |
| Devin Desktop | aggregated | native | native | native | - |
Legend:
- native - Platform auto-discovers artifacts in standard directories
- conventional - AIWG directory convention (platform reads on request)
- aggregated - Single-file compilation + discrete files for compatibility
- indexed - Full artifact bodies load through the agent’s capability search and the agent’s stable-asset loader, even without a native file loader.
Reading a behavior contract does not supply native hooks or execution support.
The indexed access audit records regression checks and remaining access restrictions.
Directory Conventions
Standard Pattern
Most providers follow `.<provider>/<type>/`:
.claude/
├── agents/ # Agent definitions
├── commands/ # Slash commands
├── skills/ # Natural language workflows
└── rules/ # Context rules
.github/
├── agents/ # Agent definitions (.agent.md format)
├── prompts/ # Slash commands (.prompt.md format)
├── instructions/ # Path-scoped rules (.instructions.md format)
└── copilot-instructions.md # Repository-wide instructionsSpecial Cases
| Provider | Special Convention |
|---|---|
| Google Antigravity CLI | Experimental, pinned to CLI 1.1.26<br>Agents → `.agents/agents/`; kernel skills → `.agents/skills/`<br>MCP → `.agents/mcp_config.json` (user scope only when explicit)<br>Global skills deferred because official paths conflict<br>Selector `antigravity`; alias and executable `agy` |
| OpenAI/Codex | Commands → `~/.codex/prompts/`<br>Skills → project `.agents/skills/` (legacy AIWG entries under `~/.codex/skills/` are pruned)<br>AGENTS.md is free-form Markdown (no YAML frontmatter or structured directives)<br>Rust CLI is current product; TypeScript CLI is legacy<br>Uses Responses API exclusively (`wire_api = "chat"` removed) |
| GitHub Copilot | Agents use `.agent.md` format<br>Commands → `.github/prompts/.prompt.md`<br>Rules → `.github/instructions/.instructions.md` (with `applyTo` globs)<br>MCP → `.vscode/mcp.json` |
| Warp Terminal | Skills natively discovered at `.warp/skills/`; agents and rules aggregated into `WARP.md`; `AGENTS.md` also supported (preferred by Warp, but `WARP.md` takes priority); `.warp/workflows/` for legacy YAML workflows |
| Devin Desktop | Agents aggregated to `AGENTS.md`<br>Commands → `.windsurf/workflows/`<br>Rules → `.windsurf/rules/*.md` (with trigger frontmatter)<br>Skills → `.windsurf/skills/`<br>Legacy selector: `windsurf` |
| Cursor | Rules use `.mdc` extension (MDC format) with frontmatter (`description`, `globs`, `alwaysApply`)<br>Skills use native `.cursor/skills/*/SKILL.md` format (2.4+)<br>Also supports `AGENTS.md` with directory inheritance<br>Legacy `.cursorrules` still generated for backward compatibility<br>Cloud Agents support MCP for remote AIWG access |
| OpenClaw | All artifacts deploy to home directory (`~/.openclaw/`)<br>First provider to support behaviors (`~/.openclaw/behaviors/`) |
| Oh My Pi | Experimental integration, distinct from Pi<br>Agents → `.omp/agents/`; prompts → `.omp/prompts/`; rules → `.omp/rules/`<br>Kernel and explicitly copied skills → one-level `.agents/skills/`; standard corpus stays lazy by default<br>Context → `.omp/AGENTS.md` native imports of WORKSPACE.md and AIWG.md<br>User scope honors OMP profiles; lifecycle bridge → `.omp/extensions/` |
| Pi Coding Agent | Commands → project `.pi/prompts/.md`<br>Portable skills and agent roles → project `.agents/skills//SKILL.md`<br>AIWG-managed standard skills → `.pi/.aiwg/skills/`<br>Context → `AGENTS.md`<br>User scope honors `${PI_CODING_AGENT_DIR:-~/.pi/agent}` |
| Hermes | Skills deploy at user scope under `~/.hermes/skills/.aiwg/`; commands and full rule bodies remain available through the agent’s capability search and the agent’s stable-asset loader |
| OpenHuman | Skills and rules deploy at user scope under `~/.openhuman/.aiwg/`; agents and commands remain available through the agent’s capability search and the agent’s stable-asset loader |
Migration Guide
Upgrading from older AIWG versions that only deployed agents?
Run the deploy command with the force option to get all four artifact types:
Use AIWG to complete this documented outcome: Run the deploy command with the force option to get all four artifact types
Have it inspect the current state, explain the plan, ask before material
changes, and report the result with verification evidence.What changes:
- New `skills/` directory created alongside `agents/`
- New `rules/` directory created alongside `agents/`
- Existing agent files remain unchanged
- Commands deployed to appropriate location per provider
No breaking changes - all existing agents remain compatible.
Agent Loop Multi-Provider Support
AIWG task loops can target different providers, not just deployment. Use the provider option to run loops through a selected adapter:
Use AIWG to complete this documented outcome: AIWG task loops can target different providers, not just deployment. Use the provider option to run loops through a selected adapter
Have it inspect the current state, explain the plan, ask before material
changes, and report the result with verification evidence.See Agent Loop Guide for full documentation.
Platform Setup Guides
| Platform | Guide |
|---|---|
| Google Antigravity CLI | Provider Guide |
| Claude Code | Setup Guide |
| OpenAI/Codex | Setup Guide |
| GitHub Copilot | Setup Guide |
| Factory AI | Setup Guide |
| Cursor | Setup Guide |
| OpenCode | Setup Guide |
| Warp Terminal | Setup Guide |
| Devin Desktop | Setup Guide |
| OpenClaw | Setup Guide |
| Grok Bot | Setup Guide |
| Hermes | Setup Guide |
| OpenHuman | Setup Guide |
| Oh My Pi | Setup Guide · Full provider guide |
| Pi Coding Agent (pi.dev) | Setup Guide |
After Setup
After deployment, follow any reload action printed by the agent-owned use operation, then ask the assistant to verify the project root, provider files it can read, installed frameworks, and one useful next action.
Try one small task immediately:
Review this project's README and getting-started docs for unclear positioning,
missing setup steps, or unsupported claims. Return the three highest-priority
fixes with file references and a recommended next edit.Success means the assistant can use the shared context in `WORKSPACE.md` and `AIWG.md`, produce a concrete review you can inspect, and name the next workflow or document to update.