The Knock Agent toolkit enables popular agent frameworks including OpenAI and Vercel's AI SDK to integrate with Knock's APIs using tools (otherwise known as function calling). It also allows you to integrate Knock into a Model Context Protocol (MCP) client such as Cursor, Windsurf, or Claude Code.
Using the Knock agent toolkit allows you to build powerful agent systems that are capable of sending cross-channel notifications to the humans who need to be in the loop. As a developer, it also helps you build Knock integrations and manage your Knock account.
You can read more in the documentation.
The Knock Agent Toolkit provides four main entry points:
@knocklabs/agent-toolkit/ai-sdk: Helpers for integrating with Vercel's AI SDK.@knocklabs/agent-tookkit/langchain: Helpers for integrating with Langchain's JS SDK.@knocklabs/agent-toolkit/openai: Helpers for integrating with the OpenAI SDK.@knocklabs/agent-toolkit/modelcontextprotocol: Low level helpers for integrating with the Model Context Protocol (MCP).
- You must have a Knock account.
- You must have a service token generated on your Knock account. The Knock agent toolkit uses your service token to interact with your Knock account.
The agent toolkit exposes a large subset of the Knock Management API and API that you might need to invoke via an agent. You can see the full list of tools in the source code.
It's possible to pass additional context to the configuration of each library to help scope the calls made by the agent toolkit to Knock. The available properties to configure are:
environment: The slug of the Knock environment you wish to execute actions in by default, such asdevelopment.userId: The user ID of the current user. When set, this will be the default passed to user tools.tenantId: The ID of the current tenant. When set, will be the default passed to any tool that accepts the tenant.
To start using the Knock MCP as a local server, you must start it with a service token. You can run it using npx.
npx -y @knocklabs/agent-toolkit -p local-mcp --service-token kst_12345
By default, the MCP server will expose all tools to the LLM. To limit the tools available you can use the --tools (-t) flag:
// Pass all tools
npx -y @knocklabs/agent-toolkit -p local-mcp --tools="*"
// Specific category
npx -y @knocklabs/agent-toolkit -p local-mcp --tools "workflows.*"
// Specific tools
npx -y @knocklabs/agent-toolkit -p local-mcp --tools "workflows.triggerWorkflow"
If you wish to enable workflows-as-tools within the MCP server, you must set the --workflows flag to pass in a list of approved workflow keys to expose. This ensures that you keep the number of tools exposed to your MCP client to a minimum.
npx -y @knocklabs/agent-toolkit -p local-mcp --workflows comment-created activate-account
It's also possible to pass environment, userId, and tenant to the local MCP server to set default values. Use the --help flag to view additional server options.
The agent toolkit provides a createKnockToolkit under the /ai-sdk path for easily integrating into the AI SDK and returning tools ready for use.
- Install the package:
npm install @knocklabs/agent-toolkit
- Import the
createKnockToolkithelper, configure it, and use it in your LLM calling:
import { createKnockToolkit } from "@knocklabs/agent-toolkit/ai-sdk";
import { openai } from "@ai-sdk/openai";
import { streamText } from "ai";
import { systemPrompt } from "@/lib/ai/prompts";
export const maxDuration = 30;
export async function POST(req: Request) {
const { messages } = await req.json();
const toolkit = await createKnockToolkit({
serviceToken: "kst_12345",
permissions: {
workflows: { read: true, run: true, manage: true },
},
});
const result = streamText({
model: openai("gpt-4o"),
messages,
tools: {
// The tools given here are determined by the `permissions`
// list above in the configuration. For instance, here we're only
// allowing the workflows
...toolkit.getAllTools(),
},
});
return result.toDataStreamResponse();
}The agent toolkit provides a createKnockToolkit under the /openai path for easily integrating into the Open AI SDK and returning tools ready for use.
- Install the package:
npm install @knocklabs/agent-toolkit
- Import the
createKnockToolkithelper, configure it, and use it in your LLM calling:
import { createKnockToolkit } from "@knocklabs/agent-toolkit/openai";
import OpenAI from "openai";
const openai = new OpenAI();
async function main() {
const toolkit = await createKnockToolkit({
serviceToken: "kst_12345",
permissions: {
// Set the permissions of the tools to expose
workflows: { read: true, run: true, manage: true },
},
});
const completion = await openai.chat.completions.create({
model: "gpt-4o",
messages,
// The tools given here are determined by the `permissions`
// list above in the configuration. For instance, here we're only
// allowing the workflows
tools: toolkit.getAllTools(),
});
// Execute the tool calls
const toolMessages = await Promise.all(
message.tool_calls.map((tc) => toolkit.handleToolCall(tc))
);
}
main();The agent toolkit provides a createKnockToolkit under the /langchain path for easily integrating into the Lanchain JS SDK and returning tools ready for use.
- Install the package:
npm install @knocklabs/agent-toolkit
- Import the
createKnockToolkithelper, configure it, and use it in your LLM calling:
import { createKnockToolkit } from "@knocklabs/agent-toolkit/langchain";
import { ChatOpenAI } from "@langchain/openai";
import { HumanMessage, SystemMessage } from "@langchain/core/messages";
import { LangChainAdapter } from "ai";
const systemPrompt = `You are a helpful assistant.`;
export const maxDuration = 30;
export async function POST(req: Request) {
const { prompt } = await req.json();
// Optional - get the auth context from the request
const authContext = await auth.protect();
// Instantiate a new Knock toolkit
const toolkit = await createKnockToolkit({
serviceToken: "kst_12345",
permissions: {
// (optional but recommended): Set the permissions of the tools to expose
workflows: { read: true, run: true, manage: true },
},
});
const model = new ChatOpenAI({ model: "gpt-4o", temperature: 0 });
const modelWithTools = model.bindTools(toolkit.getAllTools());
const messages = [new SystemMessage(systemPrompt), new HumanMessage(prompt)];
const aiMessage = await modelWithTools.invoke(messages);
messages.push(aiMessage);
for (const toolCall of aiMessage.tool_calls || []) {
// Call the selected tool by its `name`
const selectedTool = toolkit.getToolMap()[toolCall.name];
const toolMessage = await selectedTool.invoke(toolCall);
messages.push(toolMessage);
}
// To simplify the setup, this example uses the ai-sdk langchain adapter
// to stream the results back to the /langchain page.
// For more details, see: https://sdk.vercel.ai/providers/adapters/langchain
const stream = await modelWithTools.stream(messages);
return LangChainAdapter.toDataStreamResponse(stream);
}The agent toolkit provides a createKnockToolkit under the /mastra path for easily integrating into the Mastra framework and returning tools ready for use.
- Install the package:
npm install @knocklabs/agent-toolkit
- Import the
createKnockToolkithelper, configure it, and use it in your LLM calling:
import { anthropic } from "@ai-sdk/anthropic";
import { Agent } from "@mastra/core/agent";
import { Memory } from "@mastra/memory";
import { LibSQLStore } from "@mastra/libsql";
import { createKnockToolkit } from "@knocklabs/agent-toolkit/mastra";
const toolkit = await createKnockToolkit({
serviceToken: "knock_st_",
permissions: {
// (optional but recommended): Set the permissions of the tools to expose
workflows: { read: true, run: true, manage: true },
},
userId: "10",
});
export const weatherAgent = new Agent({
name: "Weather Agent",
instructions: `You are a helpful weather assistant that provides accurate weather information.`,
model: anthropic("claude-3-5-sonnet-20241022"),
tools: toolkit.getAllTools(),
memory: new Memory({
storage: new LibSQLStore({
url: "file:../mastra.db", // path is relative to the .mastra/output directory
}),
}),
});