# Harness Integrations

Cloudwright ships an MCP server (`cloudwright mcp`), a CLI, and a GitHub Action. Most AI coding
harnesses in 2026 are MCP clients, so one server reaches almost all of them, but each harness
wants the server entry in a different file, under a different JSON (or TOML) key, and each has
its own convention for the plain-text rules file that steers agent behavior. This page is the
per-harness wiring reference. Run `cloudwright integrate --list` for the same table from the
terminal, and `cloudwright integrate --harness <name>` for the exact snippet plus `--write` to
drop it straight into the right file.

Related: [MCP Reference](mcp-reference.md) | [CLI Reference](cli-reference.md) | [GitHub Action](github-action.md) | [README](../README.md)

---

## How the MCP server launches

There is no separate console script for the MCP server. Every harness below launches it the same
way: command `cloudwright`, args `["mcp"]` (the server defaults to stdio transport; pass
`["mcp", "--transport", "stdio"]` if a harness wants that spelled out explicitly). The differences
between harnesses are entirely in where that command/args pair goes and what key it sits under.

---

## Harness table

| Harness | MCP client? | Config file and key | Rules file | How to wire |
|---|---|---|---|---|
| Claude Code | Yes | `.mcp.json` (project) or `~/.claude.json` (user), key `mcpServers` | `CLAUDE.md` (does not read AGENTS.md) | `cloudwright integrate --harness claude-code --write` |
| Cursor | Yes | `.cursor/mcp.json` (project) or `~/.cursor/mcp.json`, key `mcpServers` | `AGENTS.md` (native) plus `.cursor/rules/*.mdc` | `cloudwright integrate --harness cursor --write` |
| Cline | Yes | `cline_mcp_settings.json` under the VS Code extension's global storage, key `mcpServers` | `.clinerules` (AGENTS.md support is global-only and contested) | `cloudwright integrate --harness cline --write --output <path>` |
| Windsurf | Yes | `~/.codeium/windsurf/mcp_config.json`, key `mcpServers` | `.windsurfrules` (also reads AGENTS.md) | `cloudwright integrate --harness windsurf --write` |
| GitHub Copilot | Yes | `.vscode/mcp.json`, key `servers` (not `mcpServers`), entries need `"type": "stdio"` | `.github/copilot-instructions.md` plus native AGENTS.md | `cloudwright integrate --harness copilot --write` |
| Zed | Yes | `.zed/settings.json` (project) or `~/.config/zed/settings.json`, key `context_servers` | `.rules` (falls back to AGENTS.md) | `cloudwright integrate --harness zed --write` |
| OpenAI Codex CLI | Yes | `~/.codex/config.toml`, TOML table `[mcp_servers.cloudwright]` | `AGENTS.md` (canonical origin of the convention) | `cloudwright integrate --harness codex --write` |
| JetBrains Junie | Yes | `.junie/mcp/mcp.json`, key `mcpServers` | `AGENTS.md` (primary), `.junie/guidelines.md` (legacy fallback) | `cloudwright integrate --harness junie --write` |
| Kiro | Yes | `.kiro/settings/mcp.json` (project) or `~/.kiro/settings/mcp.json`, key `mcpServers` | `.kiro/steering/*.md` (own convention); list AGENTS.md too | `cloudwright integrate --harness kiro --write` |
| Antigravity | Yes | `~/.gemini/config/mcp_config.json`, key `mcpServers` | `AGENTS.md`; confirm whether `GEMINI.md` still applies | `cloudwright integrate --harness antigravity --write` |
| Aider | No | n/a | `AGENTS.md` | See "Not an MCP client" below |

---

## Not an MCP client: Aider

Aider has no shipped MCP integration. Use the CLI pipe path instead: from an Aider session, run

```
/run cloudwright design "<description>" --json
```

and parse the structured output directly. `cloudwright integrate --harness aider` prints this
guidance along with the rules-file note (Aider reads `AGENTS.md` natively, same as most of the
harnesses above). Unofficial third-party MCP bridges for Aider exist (`mcpm-aider`,
`aider-mcp-server`) but are not something cloudwright depends on for a supported path.

---

## Discontinued or superseded (not generated by `integrate`)

`cloudwright integrate` does not generate configs for these; asking for one by name returns a
message pointing at the live successor instead.

- **Roo Code** was sunset in April 2026 and its repo was archived in May 2026. No successor is
  tracked here.
- **Continue.dev** was acquired by Cursor in June 2026; its repo is now read-only. Wire Cursor
  instead.
- **Amazon Q Developer** is being sunset by AWS in favor of **Kiro**, which carries over MCP
  support. Use `cloudwright integrate --harness kiro`.
- **Google Gemini CLI** stopped serving individual/free-tier users in June 2026 in favor of
  **Antigravity**. Use `cloudwright integrate --harness antigravity`.

---

## The rules-file block

`cloudwright integrate --rules` emits a short, harness-agnostic block that tells an agent to gate
infrastructure work through cloudwright:

```
## Cloudwright gate

Before proposing, generating, or applying cloud infrastructure or Terraform, Pulumi, or
CloudFormation changes, run cloudwright first:

- `cloudwright design "<description>" --json` for a typed architecture spec
- `cloudwright review <spec.yaml>` for an offline architecture critique (no LLM, free)
- `cloudwright cost <spec.yaml>` for a per-component cost estimate
- `cloudwright compliance <spec.yaml> --frameworks <hipaa,soc2,...>` for control-mapped findings

Treat any high or critical review or compliance finding as a blocker. Fix it before writing
Terraform, Pulumi, or CloudFormation.
```

Use `--agent-file` to pick which file it targets:

| `--agent-file` value | File | Why it exists as a separate target |
|---|---|---|
| `agents` (default) | `AGENTS.md` | Read natively by Cursor, Windsurf, Copilot, Codex, Junie, Aider, and Kiro; Zed falls back to it |
| `claude` | `CLAUDE.md` | Claude Code does not read AGENTS.md at all, so it needs its own copy |
| `gemini` | `GEMINI.md` | Gemini CLI's default context file; AGENTS.md is not read unless manually reconfigured |

`--write` appends the block to the target file (creating it if missing) behind an HTML comment
marker so a second run does not duplicate it; pass `--force` to replace an already-installed
block in place.

---

## Quick start

```bash
# See every supported harness and what it needs
cloudwright integrate --list

# Wire Claude Code and Cursor in one repo
cloudwright integrate --harness claude-code --write
cloudwright integrate --harness cursor --write

# Add the gate instructions everywhere AGENTS.md is read, plus a CLAUDE.md copy for Claude Code
cloudwright integrate --rules --write
cloudwright integrate --rules --agent-file claude --write

# Machine-readable output for scripting
cloudwright --json integrate --harness codex
```