Novu - Connect your AI agents & products to customers已验证已精选
@Novu
基本信息
分类
智能体框架
作者
Novu
关于 Novu - Connect your AI agents & products to customers
Novu is the open-source communication layer for AI agents. An agent built on Novu connects to MCP servers like Linear, GitHub, and Notion for its tools, then talks with your users on Slack, Microsoft Teams, WhatsApp, Tel


概览
What is Novu?
Novu is the open-source communication layer between AI agents and the people they serve. An agent built on Novu holds a real two-way conversation with your users on the channels they already use, and it can call any MCP server you give it while that conversation is happening.
MCP connects agents to tools. A2A connects agents to each other. ACI, Agent Communication Infrastructure, connects agents to people. Novu is that layer.
Novu never runs your agent's brain. The reasoning stays yours, in your own code or on a model platform you bring. Novu owns the rails: identity resolution, OAuth and credential management, conversation threading, tool approval, and per-channel formatting. We never run your brain. That is the whole point.
How Novu uses MCP servers
Two paths, depending on how you build the agent.
Managed Agent, no bridge server. Configure MCP servers in the Novu dashboard. Open the agent's detail page, open MCP Servers, enable the ones you want, and Novu syncs the configuration to the AI platform. The agent can then call those tools mid-conversation to look up information, create records, or act on the user's behalf.
Three authentication modes are supported:
| Mode | How it works |
|---|---|
| No auth | The server is publicly accessible. No credentials needed. |
| Novu-managed OAuth | Novu runs the OAuth flow. The end user authorizes on first use and Novu stores the tokens. |
| Provider-managed | The AI platform manages credentials in its own vault. |
For a server that needs OAuth, the agent sends the authorization link into the chat thread, the user completes the flow on the external service, and the turn resumes. Every user authorizes their own account, so the agent acts as that person and not as a shared service account.
Custom code agent, your runtime. Keep your stack and pass your own MCP tools through the framework adapter: @novu/framework/ai-sdk for the Vercel AI SDK, @novu/framework/langchain for LangChain, or the generic package for Mastra, the OpenAI SDK, the Anthropic SDK, or hand-written code. Novu calls your handler on each turn and delivers the reply.
Tool approval, in the channel the user is already in
A tool call can be paused before it runs. The agent asks the user to approve or deny it in the thread, on Slack, Microsoft Teams, WhatsApp, Telegram, or email, and the turn resumes with that decision. Human in the loop without building an approval UI.
Where the conversation happens
| Channel | Notes |
|---|---|
| Slack | Threading, typing indicators, plan cards, Block Kit, tool approval |
| Microsoft Teams | Threading, typing indicators, Adaptive Cards, tool approval |
| Session context, typing indicators, quick-reply buttons, tool approval | |
| Telegram | Threading, typing indicators, inline keyboards, edit-in-place plans, tool approval |
| Threaded HTML replies, action-link buttons, tool approval |
One conversation follows the user across channels. Identity resolution, threading, and formatting are handled per channel.
Get started
Run npx novu connect, pick a channel, sign in, and send the first message. Full walkthroughs for the AI SDK, LangChain, Mastra, and Claude managed agents are in the docs.
Open source
Novu is open source, with roughly 40,000 GitHub stars and four years of communication infrastructure behind it. Self-host the whole stack if you want to. Bring any MCP server, any model, any framework. Own the rails, never the business logic.
FAQ
Which MCP servers can I use?
Any of them. Managed agents enable servers from the dashboard, including OAuth-backed ones like Linear, GitHub, and Slack. Custom code agents pass whatever tools their framework supports. Novu does not curate your tool list.
Does Novu run the model?
No. Novu is the communication layer. The agent's intelligence is always yours, either your own code or a model platform you bring.
Who authorizes the tools, me or my users?
Your users, for anything user-scoped. Novu runs a per-subscriber OAuth flow inside the conversation and stores the tokens, so the agent acts as the person it is talking to.
Can a human approve a tool call before it runs?
Yes. Pause the turn, ask in the thread, resume with the answer. Supported on every channel above.
Where are the docs?
docs.novu.co/agents. MCP server configuration lives at docs.novu.co/agents/managed-agent/configure-mcp-servers.
智能体框架 分类下的更多 MCP 客户端
MCP Agent Graph (MAG)
keta1930MCP Agent Graph (MAG) is an agent development framework for rapidly building agent systems. This project is based on graphs, nodes, and MCP to quickly construct complex Agent systems.
MCP CLI client
adhikaspA simple CLI to run LLM prompt and implement MCP client.
MCP CLI - Model Context Protocol Command Line Interface
chrishayukMCP Client Chatbot
cgoinglove🚀 Open source MCP Client: A Multi-provider AI Chatbot Solution
Tester Client for Model Context Protocol (MCP)
apifyModel Context Protocol (MCP) Client for Apify's Actors
评论