# Install AgentWIFI

> Generated from AgentWIFI agent client registry `1.0.0`. Do not invent a client, target, path, version, endpoint, or credential flow that is not listed here.

## User request

Read https://agentwifi.ai/install.md and install AgentWIFI for this agent only.

## Quick reference

This page is the single entry point for AgentWIFI: install, doctor, and uninstall all run directly from the commands below. Nothing needs to be downloaded or saved to a file.

| Action | Command |
|---|---|
| Install | `curl -fsSL 'https://agentwifi.ai/install.sh' \| bash -s -- --target <verified-target> --version <verified-version>` |
| Doctor | `curl -fsSL 'https://agentwifi.ai/doctor.sh' \| bash -s -- --target <verified-target> --version <verified-version>` |
| Uninstall | `curl -fsSL 'https://agentwifi.ai/uninstall.sh' \| bash -s -- --target <verified-target> --version <verified-version>` |
| Windows install | `$env:AGENTWIFI_INSTALLER_TARGET='<verified-target>'; $env:AGENTWIFI_INSTALLER_VERSION='<verified-version>'; irm 'https://agentwifi.ai/install.ps1' \| iex` |
| Windows doctor | `$env:AGENTWIFI_INSTALLER_TARGET='<verified-target>'; $env:AGENTWIFI_INSTALLER_VERSION='<verified-version>'; irm 'https://agentwifi.ai/doctor.ps1' \| iex` |
| Windows uninstall | `$env:AGENTWIFI_INSTALLER_TARGET='<verified-target>'; $env:AGENTWIFI_INSTALLER_VERSION='<verified-version>'; irm 'https://agentwifi.ai/uninstall.ps1' \| iex` |

## Authorization boundary

Scope this work to AgentWIFI for the current Agent only. Start with read-only detection. Do not modify another Agent, global shell configuration, package-manager state, or unrelated configuration without asking the user first.

Never ask the user to create, copy, paste, repeat, or display an API key. Never place a credential in a command argument, URL, prompt, reply, log, screenshot, shell history, process list, fixture, or ownership manifest.

## Security and verification

The installer is distributed over HTTPS from agentwifi.ai. Before execution, every downloaded artifact (entrypoint, single-file executable, per-target manifest) is verified against the published SHA-256 release manifest. The installer uses Device authorization (it prints a verification URL and user code, and never asks for or displays a credential), writes only this agent's MCP configuration with backup, atomic write, rollback, and idempotency, then verifies the result with an MCP `account` check.

Run the exact install command for the detected target (section 1), then doctor (section 4). Expected success output: the installer prints a Device authorization URL and user code (or opens the browser), then a ready notice; doctor prints a passing MCP `account` check. If an entrypoint fails to parse or errors, re-download it fresh and retry once; if it still fails, stop and report the exact output verbatim without modifying the installer.

## 1. Detect the exact target

Identify the current Agent, operating system, and architecture without changing the environment. Continue only when the exact target appears below; if no row matches, stop and report that this target is not verified. On any platform (macOS, Linux, or Windows), first try the native MCP OAuth in section 2; Windows users who fall back to the installer use the PowerShell entrypoints in section 1.1.

| Agent | OS | Architecture | Target | Installer version |
|---|---|---|---|---|
| Codex | macos | arm64 | `codex:macos:arm64` | `1.0.0` |
| Codex | macos | x64 | `codex:macos:x64` | `1.0.0` |
| Codex | linux | arm64 | `codex:linux:arm64` | `1.0.0` |
| Codex | linux | x64 | `codex:linux:x64` | `1.0.0` |
| Codex | windows | arm64 | `codex:windows:arm64` | `1.0.0` |
| Codex | windows | x64 | `codex:windows:x64` | `1.0.0` |
| Claude Code | macos | arm64 | `claude-code:macos:arm64` | `1.0.0` |
| Claude Code | macos | x64 | `claude-code:macos:x64` | `1.0.0` |
| Claude Code | linux | arm64 | `claude-code:linux:arm64` | `1.0.0` |
| Claude Code | linux | x64 | `claude-code:linux:x64` | `1.0.0` |
| Claude Code | windows | arm64 | `claude-code:windows:arm64` | `1.0.0` |
| Claude Code | windows | x64 | `claude-code:windows:x64` | `1.0.0` |
| Cursor | macos | arm64 | `cursor:macos:arm64` | `1.0.0` |
| Cursor | macos | x64 | `cursor:macos:x64` | `1.0.0` |
| Cursor | linux | arm64 | `cursor:linux:arm64` | `1.0.0` |
| Cursor | linux | x64 | `cursor:linux:x64` | `1.0.0` |
| Cursor | windows | arm64 | `cursor:windows:arm64` | `1.0.0` |
| Cursor | windows | x64 | `cursor:windows:x64` | `1.0.0` |
| Gemini CLI | macos | arm64 | `gemini-cli:macos:arm64` | `1.0.0` |
| Gemini CLI | macos | x64 | `gemini-cli:macos:x64` | `1.0.0` |
| Gemini CLI | linux | arm64 | `gemini-cli:linux:arm64` | `1.0.0` |
| Gemini CLI | linux | x64 | `gemini-cli:linux:x64` | `1.0.0` |
| Gemini CLI | windows | arm64 | `gemini-cli:windows:arm64` | `1.0.0` |
| Gemini CLI | windows | x64 | `gemini-cli:windows:x64` | `1.0.0` |
| OpenCode | macos | arm64 | `opencode:macos:arm64` | `1.0.0` |
| OpenCode | macos | x64 | `opencode:macos:x64` | `1.0.0` |
| OpenCode | linux | arm64 | `opencode:linux:arm64` | `1.0.0` |
| OpenCode | linux | x64 | `opencode:linux:x64` | `1.0.0` |
| OpenCode | windows | arm64 | `opencode:windows:arm64` | `1.0.0` |
| OpenCode | windows | x64 | `opencode:windows:x64` | `1.0.0` |
| OpenClaw | macos | arm64 | `openclaw:macos:arm64` | `1.0.0` |
| OpenClaw | macos | x64 | `openclaw:macos:x64` | `1.0.0` |
| OpenClaw | linux | arm64 | `openclaw:linux:arm64` | `1.0.0` |
| OpenClaw | linux | x64 | `openclaw:linux:x64` | `1.0.0` |
| OpenClaw | windows | arm64 | `openclaw:windows:arm64` | `1.0.0` |
| OpenClaw | windows | x64 | `openclaw:windows:x64` | `1.0.0` |
| Hermes | macos | arm64 | `hermes:macos:arm64` | `1.0.0` |
| Hermes | macos | x64 | `hermes:macos:x64` | `1.0.0` |
| Hermes | linux | arm64 | `hermes:linux:arm64` | `1.0.0` |
| Hermes | linux | x64 | `hermes:linux:x64` | `1.0.0` |
| Hermes | windows | arm64 | `hermes:windows:arm64` | `1.0.0` |
| Hermes | windows | x64 | `hermes:windows:x64` | `1.0.0` |
| Qoder | macos | arm64 | `qoder:macos:arm64` | `1.0.0` |
| Qoder | macos | x64 | `qoder:macos:x64` | `1.0.0` |
| Qoder | linux | arm64 | `qoder:linux:arm64` | `1.0.0` |
| Qoder | linux | x64 | `qoder:linux:x64` | `1.0.0` |
| Qoder | windows | arm64 | `qoder:windows:arm64` | `1.0.0` |
| Qoder | windows | x64 | `qoder:windows:x64` | `1.0.0` |
| WorkBuddy | macos | arm64 | `workbuddy:macos:arm64` | `1.0.0` |
| WorkBuddy | macos | x64 | `workbuddy:macos:x64` | `1.0.0` |
| WorkBuddy | linux | arm64 | `workbuddy:linux:arm64` | `1.0.0` |
| WorkBuddy | linux | x64 | `workbuddy:linux:x64` | `1.0.0` |
| WorkBuddy | windows | arm64 | `workbuddy:windows:arm64` | `1.0.0` |
| WorkBuddy | windows | x64 | `workbuddy:windows:x64` | `1.0.0` |
| Kimi CLI | macos | arm64 | `kimi-cli:macos:arm64` | `1.0.0` |
| Kimi CLI | macos | x64 | `kimi-cli:macos:x64` | `1.0.0` |
| Kimi CLI | linux | arm64 | `kimi-cli:linux:arm64` | `1.0.0` |
| Kimi CLI | linux | x64 | `kimi-cli:linux:x64` | `1.0.0` |
| Kimi CLI | windows | arm64 | `kimi-cli:windows:arm64` | `1.0.0` |
| Kimi CLI | windows | x64 | `kimi-cli:windows:x64` | `1.0.0` |
| Kiro | macos | arm64 | `kiro:macos:arm64` | `1.0.0` |
| Kiro | macos | x64 | `kiro:macos:x64` | `1.0.0` |
| Kiro | linux | arm64 | `kiro:linux:arm64` | `1.0.0` |
| Kiro | linux | x64 | `kiro:linux:x64` | `1.0.0` |
| Windsurf | macos | arm64 | `windsurf:macos:arm64` | `1.0.0` |
| Windsurf | macos | x64 | `windsurf:macos:x64` | `1.0.0` |
| Windsurf | linux | arm64 | `windsurf:linux:arm64` | `1.0.0` |
| Windsurf | linux | x64 | `windsurf:linux:x64` | `1.0.0` |
| ZCode | macos | arm64 | `zcode:macos:arm64` | `1.0.0` |
| ZCode | macos | x64 | `zcode:macos:x64` | `1.0.0` |
| ZCode | linux | arm64 | `zcode:linux:arm64` | `1.0.0` |
| ZCode | linux | x64 | `zcode:linux:x64` | `1.0.0` |

## 1.1 Windows (verified PowerShell entrypoints)

If the host is Windows and your Agent is listed below, fall back here to the PowerShell entrypoint (instead of the bash command) only after the native OAuth path in section 2 is unavailable or failed. The PowerShell entrypoints run directly on Windows PowerShell 5.1 or newer and do not require a Node.js runtime; the shell downloads the AgentWIFI installer executable, verifies its SHA-256 checksum, and then runs it.


Before running the entrypoints, set `AGENTWIFI_INSTALLER_TARGET` and `AGENTWIFI_INSTALLER_VERSION` to the exact target and version from the table (PowerShell entrypoints cannot receive arguments through `irm | iex`).

```powershell
$env:AGENTWIFI_INSTALLER_TARGET = '<verified-target>'
$env:AGENTWIFI_INSTALLER_VERSION = '<verified-version>'
irm 'https://agentwifi.ai/install.ps1' | iex
irm 'https://agentwifi.ai/doctor.ps1' | iex
irm 'https://agentwifi.ai/uninstall.ps1' | iex
```

## 2. Preferred path: native MCP OAuth (no installer)

For these agents, prefer the agent's own MCP OAuth login instead of downloading the installer. Run the exact add command for the detected agent, then complete authorization in the browser that opens. The same commands apply on macOS, Linux, and Windows. The agent registers itself and stores its own token; no AgentWIFI installer download is needed.

| Agent | Add MCP server | Complete authorization |
|---|---|---|
| Codex | `codex mcp add agentwifi --url https://mcp.agentwifi.ai/v1/mcp` | `codex mcp login agentwifi` |
| Claude Code | `claude mcp add --transport http agentwifi https://mcp.agentwifi.ai/v1/mcp` | Open /mcp and choose Authenticate for agentwifi, then complete browser authorization. |
| Gemini CLI | `gemini mcp add agentwifi --scope user --transport http https://mcp.agentwifi.ai/v1/mcp` | Complete authorization in the browser that opens. |
| Kimi CLI | `kimi mcp add --transport http --auth oauth agentwifi https://mcp.agentwifi.ai/v1/mcp` | `kimi mcp auth agentwifi` |

Before completing OAuth login, check whether this agent already has an AgentWIFI connection created by the device-flow installer. That connection stores an `Authorization: Bearer aw...` header inside the existing `agentwifi` MCP entry, and the header wins over OAuth: the server keeps serving the old account's balance and tools. If such a header exists (for example, inspect the `agentwifi` entry of the agent's own MCP config, such as `codex mcp get agentwifi`), remove the header or delete and re-add the server entry before running the login command, then complete OAuth authorization. Never print or repeat the header value.

After the login command reports success, keep the reply short and skip the intermediate narration (reading install.md, detecting the environment, running add/login, or the current-session MCP server state). Tell the user only (1) AgentWIFI is connected and the next step is to open a new session or reload the agent to load the tools, and (2) a few ready-to-use example prompts, such as "search the latest AI news", "read the key points of <url>", and "get the latest stock price for AAPL".

Verify the OAuth install with the agent's own MCP list command (for example `codex mcp list`). Do not run `doctor.sh` here: it only verifies the Device Flow installer and will report an unmanaged or incomplete managed config for an OAuth connection.

For Codex specifically, if it runs on a custom model catalog (for example DeepSeek via `model_catalog_json`), the catalog's `supports_search_tool: true` can make MCP tools invisible because third-party providers do not implement `tool_search`. If `codex mcp list` shows the server but the tools never appear, set `supports_search_tool` to `false` for each entry in `~/.codex/models.json` (back up the file first), then reload Codex or open a new session. Running `doctor.sh` for Codex applies this correction automatically.

If the current agent is not listed above, or the OAuth flow fails, continue to the device-flow installer below.

## 3. Fallback: device-flow installer

Run the public installer with the exact target and version from the table:

```sh
curl -fsSL 'https://agentwifi.ai/install.sh' | bash -s -- --target <verified-target> --version <verified-version>
```

The installer opens the AgentWIFI Device authorization page. If it cannot open a browser, show the verification URL and user code exactly as printed. Let the installer poll at the server-provided interval. Do not replace Device Flow with a manual credential prompt.

The installer reads existing configuration first, merges only the `agentwifi` entry, creates a backup before mutation, writes atomically, records ownership, rolls back on failure, and is safe to repeat. Do not edit JSON, TOML, or YAML by hand instead of running it.

## 4. Verify, recover, or uninstall

After installation, run doctor with the same exact target and version. Report success only when doctor completes the real MCP `account` check:

```sh
curl -fsSL 'https://agentwifi.ai/doctor.sh' | bash -s -- --target <verified-target> --version <verified-version>
```

If installation or doctor fails, preserve the reported backup and managed files, do not expose credential-bearing configuration, and stop. To uninstall, remove only unchanged manifest-owned AgentWIFI content:

```sh
curl -fsSL 'https://agentwifi.ai/uninstall.sh' | bash -s -- --target <verified-target> --version <verified-version>
```

## 4.1 After script installation, tell the user the exact next step

After the installer reports success, do not claim the tools are already visible in the current session. Clients load MCP configuration when they start, so tell the user:

1. Configuration is written and verified; no credential entry is needed.
2. For WorkBuddy only, after browser authorization open WorkBuddy, click 「Experts·Skills·Connectors」 (专家·技能·连接器) in the top-left menu, open Connectors (连接器), click Custom Connectors (自定义连接器) at the top right, then find AgentWIFI (the installer already added it) and click Trust (信任). This is a required manual confirmation in WorkBuddy; do not claim the tools are ready or try to approve the connector for the user.
3. Open a new session in the installed agent (or reload it) so AgentWIFI tools can appear.
4. After the session opens, try one of the example prompts from the installer output.

## 5. Restricted fallback when shell execution is unavailable

Do not improvise direct MCP configuration. A no-shell Agent may install only the exact verified AgentWIFI Skill target below by reading the hosted manifest and its listed files. It must preserve unowned or modified content, use a temporary sibling plus atomic rename, record ownership, remain idempotent, and roll back every file on failure. After installation, read only that Skill's metadata and referenced resources.

| Agent | OS | Architecture | Verified Skill manifest | Install root | Ownership manifest |
|---|---|---|---|---|---|
| Codex | macos | arm64 | https://agentwifi.ai/skills/agentwifi/manifest.json | `.agents/skills/agentwifi` | `.agents/skills/.agentwifi-agentwifi-manifest.json` |

If the table has no exact match, or the Agent cannot verify the manifest and perform ownership-safe writes, stop. Installing the Skill alone does not prove MCP configuration. Until the Agent has a verified secret-safe Device Flow and configuration adapter, ask the user to use the public installer instead; never fall back to exposing a credential in chat.

## 5.1 After installing the Prompt Skill, explain how to use it now

After the Prompt Skill is installed at the listed install root, do not tell the user that restarting is the only way to use it. The Skill is an instruction file on disk, so the current session can follow it immediately:

1. State the exact installed path, for example `<project-root>/.agents/skills/agentwifi/SKILL.md`.
2. Tell the user the current session can already read that SKILL.md and follow its setup guidance; AgentWIFI MCP tools appear after a new task opens (or the client reloads).
3. Tell the user that a new task opened in the AgentWIFI project will automatically list the Skill in its skill directory.
4. Do not claim MCP configuration succeeded unless the verified Device Flow and doctor account check completed.

## 6. After installation, show the user what AgentWIFI can do

After doctor passes, fetch the actual capability list from the configured AgentWIFI MCP connection with `list_tools` (or `find_tools` for keyword search) before claiming anything. Then give the user:

1. A short, grouped summary of the available capability families (for example web search/read, social posts and profiles, ecommerce products, media, finance) — only the ones `list_tools` actually returned.
2. Three to five ready-to-use example prompts from common scenarios, written as natural-language requests the user can copy into the agent chat, matching the tools' real input contracts.

Keep the summary brief. Never advertise tools outside the current catalog, and never fabricate IDs, URLs, or capabilities that did not come from `list_tools`.

Canonical guide: https://agentwifi.ai/install.md
