Replai
SaaS製品向けのAIカスタマーサポートエージェントをセットアップ、テスト、公開します。
ホスト型 MCP サーバー
npx add-mcp 'https://api.replai.tw/mcp'Claude Code、Codex、Cursor などにインストールできます
ドキュメント
Connect Replai to Claude Code, Codex, ChatGPT or another MCP client
Replai’s remote MCP server is https://api.replai.tw/mcp. Claude and ChatGPT add it by URL, Cursor and VS Code install it with one click, and Claude Code, Codex and Grok Build add it with one command. The first connection opens your browser to sign in, choose one Workspace and approve access, and you can revoke that access at any time.
Before you start
You need a Replai account, an AI tool that supports remote MCP servers with OAuth sign-in, and a browser on the same computer to approve access. No API key or token is copied by hand.
- Replai account: sign in with Google at replai.tw. Replai creates a Workspace the first time you sign in.
- AI tool: Claude, ChatGPT, Cursor, VS Code, Claude Code, Codex, Grok Build or another client that connects to remote MCP servers over Streamable HTTP and signs in with OAuth.
- Browser: the first connection opens Replai in your browser, where you approve access.
Connect Claude
Add Replai as a custom connector by its URL, then connect it. This works on claude.ai and in the Claude apps on every plan; the Free plan allows one custom connector.
https://api.replai.tw/mcp
- In Claude, open Customize → Connectors and select Add custom connector.
- Enter the URL above and select Add. Claude signs in with its published client identity, so no client ID or secret is needed.
- Select Connect next to Replai and approve access in the window that opens.
On Team and Enterprise plans an Owner adds the connector under Organization settings → Connectors, and each member then connects with their own Replai account.
Connect Claude Code
Add Replai as a remote HTTP server, then authenticate it from inside Claude Code.
claude mcp add --transport http replai https://api.replai.tw/mcp
- Run the command above in a terminal.
- Start Claude Code and run /mcp, choose replai and select Authenticate.
- Approve access in the browser window that opens, then return to Claude Code.
Connect Codex
Add Replai by URL. Codex discovers Replai’s sign-in on its own, so no further settings are needed.
codex mcp add replai --url https://api.replai.tw/mcp
- Run the command above in a terminal.
- Approve access in the browser window that opens. If no window opens, run codex mcp login replai.
- Start Codex and ask it to work in Replai.
Connect Grok Build
Add Replai as a remote HTTP server, then sign in from the MCP list inside Grok.
grok mcp add --transport http replai https://api.replai.tw/mcp
- Run the command above in a terminal.
- Start Grok, run /mcps, choose replai and press i.
- Approve access in the browser window that opens, then return to Grok.
Connect ChatGPT
Add Replai as a custom MCP server in ChatGPT, where your plan and workspace admin allow it.
https://api.replai.tw/mcp
- Open ChatGPT Plugins at chatgpt.com/plugins, select + and choose Add custom MCP server.
- Name it Replai, choose Public endpoint, enter the URL above and choose OAuth as the authentication.
- Accept the warning, select Create as a plugin, then approve access in the Replai window that opens.
Connect Cursor
Install Replai in Cursor with one click, or add it to your Cursor MCP settings by hand.
- Select Add Replai to Cursor above and confirm the install in Cursor.
- When Cursor connects, approve access in the browser window that opens.
To add it by hand, put this server in ~/.cursor/mcp.json:
{ "mcpServers": { "replai": { "url": "https://api.replai.tw/mcp" } } }
Connect VS Code
Install Replai in VS Code with one click, or add it from a terminal.
code --add-mcp '{"name":"replai","type":"http","url":"https://api.replai.tw/mcp"}'
- Select Add Replai to VS Code above, or run the command in a terminal.
- Start the replai server when VS Code asks, then approve access in the browser window that opens.
- Use Replai’s tools from Copilot Chat in agent mode.
Connect another MCP client
Any client that connects to remote MCP servers uses the same URL and signs in with OAuth.
https://api.replai.tw/mcp
- Transport: Streamable HTTP.
- Authentication: OAuth with PKCE. Replai publishes its sign-in details for clients to discover and accepts a client metadata document URL as the client ID, the method Claude and ChatGPT use; dynamic client registration also works.
What the first connection asks for
Your browser opens Replai and asks you to sign in with your Replai account. You then choose the one Workspace this AI tool can reach, review what it will be able to do and select Allow access. The browser returns you to the AI tool.
- Sign in with Google, using the account you use for Replai.
- Choose the Workspace the AI tool can access.
- Review the permissions the AI tool asks for, then select Allow access.
- Return to your AI tool. It can now use Replai’s tools in that Workspace.
Every connected AI tool can read your Agents, knowledge, tests, conversations and Workspace information. Replai also asks for permission to edit Agent drafts, create Agents, publish, manage knowledge sources, tests and channels, reply in the Inbox, manage linked engineering issues, connect external resources, update Workspace settings or remove Agents; the sign-in page lists each permission the tool receives. Edits to an Agent stay in its draft until the Agent is published.
What your AI tool asks before it acts
Reads run without asking. Every tool that changes something is labelled as a change, and tools that message a customer, publish, overwrite settings or remove something are labelled as potentially destructive, so Claude, ChatGPT and other clients ask you before each of those calls unless you choose to always allow it.
- Messages to customers: replying, approving a draft, taking over a conversation or returning it to the AI sends a message the customer sees. It cannot be unsent.
- Publishing: edits stay in the Agent’s draft. They go live only when publication_publish publishes the exact candidate publication_review showed.
- Removing: removing an Agent, a channel or a knowledge source first reads its impact. The removal is refused if anything changed after that check.
- Permissions: each kind of change has its own permission, and the sign-in page lists them all. A tool connected with fewer permissions is refused a change it was not allowed until you reconnect and approve it.
Which tools Replai provides
The tools cover the whole setup path and daily support work: Agents and their drafts, knowledge, tests, publication, channels, the Inbox, engineering issues and analytics. Your AI tool reads the exact list from Replai when it connects.
Replai MCP tools by area
| Area | What your AI tool can do | Example tools |
|---|---|---|
| Workspace and Agents | See the connected Workspace, list and create Agents, read and edit an Agent’s draft | connection_context, agents_list, agent_create, agent_draft_edit |
| Knowledge | Discover and import website pages, import Google Drive files and GitHub repositories, add text, refresh or withdraw sources | website_discover, website_import, source_add, source_refresh |
| Tests and gaps | Save questions as tests, run them, read results and review unanswered questions | test_case_create, test_case_run, knowledge_gaps_list |
| Publication | Review the exact candidate, publish it, pause, or restore an earlier version into the draft | publication_review, publication_publish, publication_restore |
| Channels | Turn on the website widget and read its install code, adjust widget and LINE group settings, check LINE health | widget_enable, widget_installation_read, line_behavior_edit |
| Inbox | Read conversations, reply, approve or reject drafts, take over, return to AI and resolve | inbox_list, inbox_reply, inbox_draft_approve, inbox_return_to_ai |
| Engineering | Propose and file Linear issues from conversations, record releases and track customer follow-ups | engineering_propose, engineering_approve, engineering_release_decide |
| Insights | Read Agent analytics, channel activity and usage | agent_analytics, agent_kpis_read, workspace_usage |
A few steps stay in the browser by design. Uploading a file from your computer, entering LINE channel credentials and connecting Google Drive, GitHub or Linear each return a Replai link for you to open; the AI tool then reads the result.
Revoke access
Open Workspace settings → MCP & CLI. Under Connected, select Revoke access next to the tool. It loses access immediately.
- In Replai, open Workspace settings → MCP & CLI.
- Find the AI tool under Connected. Each entry shows its permissions and when its access expires.
- Select Revoke access.
To also remove Replai from the AI tool, run claude mcp remove replai in Claude Code, codex mcp remove replai in Codex or grok mcp remove replai in Grok Build. Removing it there alone does not revoke the access in Replai. In Claude or ChatGPT, remove the connector from the same settings page where you added it.
Use the Replai CLI instead
The Replai CLI runs the same setup, test, publication and Inbox work from a terminal, which suits coding agents that work through shell commands. Install it, then sign in once in the browser.
curl -fsSL https://replai.tw/install.sh | sh
- Run the install command above.
- Run replai login and approve access in the browser.
- Run replai status to see the connected Workspace, and replai --help for the commands.
Commands cover Agents, drafts, knowledge sources, website import, the widget, tests, publication review and publishing, and the Inbox. Results print as JSON. Run replai logout to revoke the CLI’s access and remove its saved credentials.
Changes
Replai’s MCP server keeps one version across its Server Card, the MCP Registry and the tools your AI tool sees. The newest row is the version running now.
Replai MCP server changes
| Version | Date | What changed |
|---|---|---|
| 1.12.0 | Oct 29, 2026 | A publication review now shows the welcome messages in every language and any deleted knowledge source the release removes, exactly as publishing will apply them. A refused publish says whether a channel or payment setup is missing. |
| 1.11.3 | Oct 28, 2026 | The review test for a single preview run now checks what the run guarantees: one saved question, one run, and an honest report of the answer. |
| 1.11.2 | Oct 28, 2026 | Publishing and pausing an Agent are now labeled as actions that reach your customers’ channels. |
| 1.11.1 | Oct 28, 2026 | The plugin listing now opens with what Replai does and says when a change reaches your customers. |
| 1.11.0 | Oct 28, 2026 | Before publishing, Replai now asks for payment setup instead of a subscription. Test runs and knowledge imports count toward a monthly setup allowance; when it runs out, the tool says when it resets. |
| 1.10.0 | Oct 28, 2026 | Connecting asks for every permission at once, so the AI tool can make changes without signing in again. A tool connected earlier gets them when it reconnects. The setup skill names Workspace settings → MCP & CLI for revoking access. |
| 1.9.1 | Oct 27, 2026 | The Agent KPI and analytics tools say which one carries resolution and handoff figures. |
| 1.9.0 | Oct 27, 2026 | Tool descriptions state what each tool does, and the server explains the shared workflow to your AI tool. Claude and ChatGPT sign in with their published client identity, and a connection unused for 30 days expires; reconnect to continue. |
| 1.8.0 | Oct 26, 2026 | Connecting Google Drive or Gmail from your AI tool reports its progress, like the other providers. |
| 1.4.0–1.7.0 | Oct 26, 2026 | Connecting GitHub, Linear, Meta, Discord or LINE from your AI tool returns a link to open and a status to follow. |
| 1.3.0 | Oct 25, 2026 | Discord channels. |
| 1.2.0 | Oct 25, 2026 | Every tool returns a typed result, and tool titles and effect labels were reviewed. |
| 1.1.0 | Oct 25, 2026 | Reading a conversation’s messages no longer changes the conversation. |
| 1.0.0 | Oct 20, 2026 | First version listed in the official MCP Registry. |
Set up a support Agent from Claude Code or Codex
Keep reading
- How Replai works: Replai answers customers from your product knowledge and hands off to a person when needed. Setup steps, answer settings and ways to improve answers.
- Knowledge sources: Files, web pages, websites, Google Drive, GitHub repositories and text: what a Replai Agent can learn from, the limits of each source, and how to keep knowledge current.
- Channels: One Agent answers on your website or in your app, a LINE Official Account including group chats, Gmail, Messenger and Instagram, with the same knowledge and handoff settings. WhatsApp status included.
Replai · AIthentic LLC Last updated October 27, 2026