SDK

screenpipe provides two sdk packages:

@screenpipe/js - for node.js environments (nextjs api routes, etc)
@screenpipe/browser - for browser environments

both sdks provide type-safe interfaces to interact with screenpipe’s core functionality. feel free to use our docs as context in cursor agent through MCP

installation

node.js sdk

npm
pnpm
bun
yarn

npm install @screenpipe/js

browser sdk

npm
pnpm
bun
yarn

npm install @screenpipe/browser

basic usage

// node.js
import { pipe } from '@screenpipe/js'

// browser
import { pipe } from '@screenpipe/browser'

search api

const results = await pipe.queryScreenpipe({
  q: "john",
  contentType: "ocr", // "ocr" | "audio" | "ui" | "all" | "audio+ui" | "ocr+ui" | "audio+ocr"
  limit: 10,
  offset: 0,
  startTime: "2024-03-10T12:00:00Z",
  endTime: "2024-03-10T13:00:00Z",
  appName: "chrome",
  windowName: "meeting",
  includeFrames: true,
  minLength: 10,
  maxLength: 1000,
  speakerIds: [1, 2],
  frameName: "screenshot.png",
  browserUrl: "github.com/mediar-ai/screenpipe" // Filter by browser URL
})

common usage patterns

fetching recent screen activity

// Get the last hour of screen activity from Chrome
const oneHourAgo = new Date(Date.now() - 60 * 60 * 1000).toISOString()
const now = new Date().toISOString()

const results = await pipe.queryScreenpipe({
  contentType: "ocr",
  startTime: oneHourAgo,
  endTime: now,
  appName: "chrome",
  limit: 50,
  includeFrames: true, // include base64 encoded images
})

// Process results
for (const item of results.data) {
  console.log(`At ${item.content.timestamp}:`)
  console.log(`Text: ${item.content.text}`)
  
  if (item.content.frame) {
    // You could display or process the base64 image
    // e.g., <img src={`data:image/png;base64,${item.content.frame}`} />
  }
}

searching for specific content

const results = await pipe.queryScreenpipe({
  browserUrl: "discord*",
  startTime: new Date(Date.now() - 1000 * 60 * 60 * 24).toISOString(),
  limit: 20,
})

// Process results based on content type, when using browserUrl, only OCR is available
for (const item of results.data) {
  if (item.type === "OCR") {
    console.log(`Screen text: ${item.content.text}`)
  } else if (item.type === "Audio") {
    console.log(`Audio transcript: ${item.content.transcription}`)
    console.log(`Speaker: ${item.content.speaker_id}`)
  } else if (item.type === "UI") {
    console.log(`UI element: ${item.content.element_type}`)
  }
}

building a timeline of user activity

// Get a full day of activity
const startOfDay = new Date()
startOfDay.setHours(0, 0, 0, 0)

const endOfDay = new Date() 
endOfDay.setHours(23, 59, 59, 999)

// Get both screen and audio data
const results = await pipe.queryScreenpipe({
  contentType: "ocr+audio",
  startTime: startOfDay.toISOString(),
  endTime: endOfDay.toISOString(),
  limit: 1000, // Get a large sample
})

// Group by hour for timeline visualization
const timelineByHour = {}

for (const item of results.data) {
  const timestamp = new Date(item.content.timestamp)
  const hour = timestamp.getHours()
  
  // Initialize hour bucket if needed
  if (!timelineByHour[hour]) {
    timelineByHour[hour] = {
      screenEvents: 0,
      audioEvents: 0,
      apps: new Set(),
    }
  }
  
  // Update counts
  if (item.type === "OCR") {
    timelineByHour[hour].screenEvents++
    
    if (item.content.app_name) {
      timelineByHour[hour].apps.add(item.content.app_name)
    }
  } else if (item.type === "Audio") {
    timelineByHour[hour].audioEvents++
  }
}

// Now timelineByHour can be used to build a visualization
console.log(timelineByHour)

fetching website-specific content

/**
 * Retrieves all content captured from a specific website
 * This is useful for aggregating research materials or tracking 
 * activity on particular web applications
 */
async function getWebsiteCaptures(domain: string, days = 7) {
  const startTime = new Date(Date.now() - days * 24 * 60 * 60 * 1000).toISOString();
  const endTime = new Date().toISOString();

  const results = await pipe.queryScreenpipe({
    browserUrl: domain,  // Target specific website
    contentType: "ocr",  // Get both visual content and UI elements
    startTime,
    endTime,
    limit: 100,
    includeFrames: true  // Include screenshots
  });

  // Process and analyze the results
  const captures = results.data.map(item => {
    const timestamp = new Date(
      item.type === "OCR" ? item.content.timestamp : item.content.timestamp
    );
    
    return {
      timestamp,
      date: timestamp.toLocaleDateString(),
      time: timestamp.toLocaleTimeString(),
      type: item.type,
      content: item.type === "OCR" ? item.content.text : item.content.text,
      screenshot: item.type === "OCR" && item.content.frame ? item.content.frame : null,
      url: item.type === "OCR" ? item.content.browserUrl : item.content.browserUrl
    };
  });

  return captures;
}

// Example: Get all GitHub activity from the past week
const githubActivity = await getWebsiteCaptures("github.com");
console.log(`Found ${githubActivity.length} interactions with GitHub`);

// Example: Search for specific text within a website
const searchTerm = "pull request";
const pullRequests = githubActivity.filter(item => 
  item.content.toLowerCase().includes(searchTerm.toLowerCase())
);
console.log(`Found ${pullRequests.length} pull request interactions`);

vercel-like crons

you need to add a pipe.json file to your pipe folder with this config for example:

{
  "crons": [
    {
      "path": "/api/log",
      "schedule": "0 */5 * * *" // every 5 minutes
    }
  ]
}

this will run the /api/log route every 5 minutes. check how the obsidian pipe implements this, route and pipe.json for a complete example. we recommend using the CLI to add the update-pipe-config server action to your pipe. this will allow you to update the pipe’s cron schedule using a server action.

bunx @screenpipe/dev@latest components add
# select "update-pipe-config" from the menu

please adjust its code to your needs as some things are hardcoded in it.

realtime streams

// stream transcriptions
for await (const chunk of pipe.streamTranscriptions()) {
  console.log(chunk.choices[0].text)
  console.log(chunk.metadata) // { timestamp, device, isInput }
}

// stream vision events
for await (const event of pipe.streamVision(true)) { // true to include images
  console.log(event.data.text)
  console.log(event.data.app_name)
  console.log(event.data.image) // base64 if includeImages=true
}

react hooks sdk support ⚛️

screenpipe provides first-class support for React applications through custom hooks, enabling seamless integration with your React components. while you can manually create hooks using libraries like React Query, we recommend leveraging our built-in CLI to quickly add pre-built, optimized hooks to your pipes.

using the cli to add hooks

the fastest way to integrate react hooks into your pipe is through our CLI:

bunx --bun @screenpipe/dev@latest components add

select from the interactive menu to add hooks such as:

use-pipe-settings: manage pipe-specific and global app settings.
use-health: monitor pipe health and status.
use-ai-provider: integrate seamlessly with ai providers.
use-sql-autocomplete: provide sql query assistance.

these hooks follow best practices, ensuring type safety, efficient state management, and easy integration with your existing React components.

real-time meeting summarizer

⚠️ make sure to enable real time transcription in the screenpipe app settings or CLI args.

import OpenAI from 'openai'

const openai = new OpenAI({
  apiKey: process.env.OPENAI_API_KEY,
})

async function summarizeMeeting() {
  let transcript = ""
  let summaryBuffer = ""
  let lastSummaryTime = Date.now()
  
  // Stream transcriptions in real-time
  for await (const chunk of pipe.streamTranscriptions()) {
    const text = chunk.choices[0].text
    transcript += text + " "
    
    // Generate summary every 30 seconds
    if (Date.now() - lastSummaryTime > 30000 && transcript.length > 200) {
      const summary = await openai.chat.completions.create({
        model: "gpt-3.5-turbo",
        messages: [
          {
            role: "system",
            content: "You summarize meeting transcripts concisely."
          },
          {
            role: "user",
            content: `Summarize this meeting transcript: ${transcript}`
          }
        ]
      })
      
      summaryBuffer = summary.choices[0].message.content
      lastSummaryTime = Date.now()
      
      // Display or store the summary
      console.log("Meeting Summary Update:", summaryBuffer)
    }
  }
}

// Start summarizing
summarizeMeeting().catch(console.error)

real-time screen activity monitor

async function monitorScreenActivity() {
  // Monitor screen content in real-time
  for await (const event of pipe.streamVision(true)) {
    // Check if we're looking at important apps
    const appName = event.data.app_name?.toLowerCase() || ""
    const text = event.data.text?.toLowerCase() || ""
    
    // Example: Detect when looking at email
    if (
      appName.includes("outlook") || 
      appName.includes("gmail") ||
      text.includes("inbox") ||
      text.includes("compose email")
    ) {
      console.log("Email activity detected at", new Date().toLocaleTimeString())
      
      // You could log this, send notifications, etc.
    }
    
    // Example: Track time spent in different applications
    // This would require keeping state between events
  }
}

// Start monitoring
monitorScreenActivity().catch(console.error)

settings management

pipes can access and modify screenpipe app settings through the SDK. this is useful for storing pipe-specific configuration and accessing global app settings.

quick start with CLI

the fastest way to add settings management to your pipe is using our CLI:

bunx --bun @screenpipe/dev@latest components add
# select "use-pipe-settings" from the menu

this will add the following components to your pipe:

use-pipe-settings hook for react components
get-screenpipe-app-settings server action
required typescript types

manual setup

create types for your settings:

// src/lib/types.ts
import type { Settings as ScreenpipeAppSettings } from '@screenpipe/js'

export interface Settings {
  // your pipe specific settings
  customSetting?: string
  anotherSetting?: number
  
  // screenpipe app settings
  screenpipeAppSettings?: ScreenpipeAppSettings
}

create server action to access settings:

// src/lib/actions/get-screenpipe-app-settings.ts
import { pipe } from '@screenpipe/js'
import type { Settings as ScreenpipeAppSettings } from '@screenpipe/js'

export async function getScreenpipeAppSettings() {
  return await pipe.settings.getAll()
}

export async function updateScreenpipeAppSettings(
  newSettings: Partial<ScreenpipeAppSettings>
) {
  return await pipe.settings.update(newSettings)
}

create react hook for settings management:

// src/lib/hooks/use-pipe-settings.tsx
import { useState, useEffect } from 'react'
import { Settings } from '@/lib/types'
import {
  getScreenpipeAppSettings,
  updateScreenpipeAppSettings,
} from '@/lib/actions/get-screenpipe-app-settings'

const DEFAULT_SETTINGS: Partial<Settings> = {
  customSetting: 'default value',
  anotherSetting: 42,
}

export function usePipeSettings() {
  const [settings, setSettings] = useState<Partial<Settings> | null>(null)
  const [loading, setLoading] = useState(true)

  useEffect(() => {
    loadSettings()
  }, [])

  const loadSettings = async () => {
    try {
      // load screenpipe app settings
      const screenpipeSettings = await getScreenpipeAppSettings()

      // get pipe specific settings from customSettings
      const pipeSettings = {
        ...(screenpipeSettings.customSettings?.yourPipeName && {
          ...screenpipeSettings.customSettings.yourPipeName,
        }),
      }

      // merge everything together
      setSettings({
        ...DEFAULT_SETTINGS,
        ...pipeSettings,
        screenpipeAppSettings: screenpipeSettings,
      })
    } catch (error) {
      console.error('failed to load settings:', error)
    } finally {
      setLoading(false)
    }
  }

  const updateSettings = async (newSettings: Partial<Settings>) => {
    try {
      // split settings
      const { screenpipeAppSettings, ...pipeSettings } = newSettings

      const mergedPipeSettings = {
        ...DEFAULT_SETTINGS,
        ...pipeSettings,
      }

      // update screenpipe settings if provided
      await updateScreenpipeAppSettings({
        ...screenpipeAppSettings,
        customSettings: {
          ...screenpipeAppSettings?.customSettings,
          yourPipeName: pipeSettings,
        },
      })

      // update state with everything
      setSettings({
        ...mergedPipeSettings,
        screenpipeAppSettings:
          screenpipeAppSettings || settings?.screenpipeAppSettings,
      })
      return true
    } catch (error) {
      console.error('failed to update settings:', error)
      return false
    }
  }

  return { settings, updateSettings, loading }
}

use in your components:

import { usePipeSettings } from '@/lib/hooks/use-pipe-settings'

export function SettingsComponent() {
  const { settings, updateSettings, loading } = usePipeSettings()

  if (loading) return <div>loading...</div>

  return (
    <form onSubmit={async (e) => {
      e.preventDefault()
      const formData = new FormData(e.target as HTMLFormElement)
      await updateSettings({
        customSetting: formData.get('customSetting') as string,
        anotherSetting: parseInt(formData.get('anotherSetting') as string),
      })
    }}>
      <input 
        name="customSetting"
        defaultValue={settings?.customSetting}
      />
      <input 
        name="anotherSetting"
        type="number"
        defaultValue={settings?.anotherSetting}
      />
      <button type="submit">save</button>
    </form>
  )
}

best practices

store pipe-specific settings under customSettings.yourPipeName in screenpipe app settings
use typescript for type safety
provide default values for all settings
handle loading and error states
validate settings before saving
use server actions for settings operations
consider using shadcn/ui components for consistent UI

see the obsidian pipe for a complete example of settings management.

integrating with llms

screenpipe’s sdk can be easily integrated with various ai providers to analyze and generate insights from screen activity. here are common patterns for connecting context data with llms. if the user is using screenpipe-cloud, you can use claude-3-7-sonnet, gemini-2.0-flash-lite or gpt-4o (Anthropic, Google, OpenAI or local models).

vercel ai sdk integration

the vercel ai sdk offers a streamlined way to work with llms. this example from the obsidian pipe demonstrates how to generate structured work logs:

import { generateObject, embed } from "ai";
import { ollama } from "ollama-ai-provider";
import { pipe } from "@screenpipe/js";
import { z } from "zod";

// define your output schema
const workLog = z.object({
  title: z.string(),
  description: z.string(),
  tags: z.array(z.string()),
  mediaLinks: z.array(z.string()).optional(),
});

type WorkLog = z.infer<typeof workLog> & {
  startTime: string;
  endTime: string;
};

async function generateWorkLog(screenData, model, startTime, endTime) {
  // configure the prompt with context and instructions
  const prompt = `You are analyzing screen recording data from Screenpipe.
    Based on the following screen data, generate a concise work activity log entry.
    
    Screen data: ${JSON.stringify(screenData)}
    
    Return a JSON object with:
    {
        "title": "Brief title describing the main activity",
        "description": "Clear description of what was accomplished",
        "tags": ["#relevant-tool", "#activity-type"],
        "mediaLinks": ["<video src=\"file:///path/to/video.mp4\" controls></video>"]
    }`;

  // use ollama provider with vercel ai sdk
  const provider = ollama(model);
  const response = await generateObject({
    model: provider,
    messages: [{ role: "user", content: prompt }],
    schema: workLog, // zod schema for type safety
  });

  return {
    ...response.object,
    startTime: formatDate(startTime),
    endTime: formatDate(endTime),
  };
}

// api endpoint implementation
async function handleRequest() {
  // get last hour of activity
  const now = new Date();
  const oneHourAgo = new Date(now.getTime() - 3600000);

  // query context from screenpipe
  const screenData = await pipe.queryScreenpipe({
    startTime: oneHourAgo.toISOString(),
    endTime: now.toISOString(),
    limit: 100,
    contentType: "all",
  });

  // generate structured log with ai
  const logEntry = await generateWorkLog(
    screenData.data,
    "llama3.2", // ollama model name
    oneHourAgo,
    now
  );

  return { message: "log generated successfully", logEntry };
}

see the obsidian pipe for the complete implementation. or irs-agent for a more complex example.

deduplicating context with embeddings

when working with large amounts of screen data, it’s useful to remove duplicates before sending to an llm:

import { embed } from "ai";
import { ollama } from "ollama-ai-provider";
import { ContentItem } from "@screenpipe/js";

async function deduplicateScreenData(screenData: ContentItem[]): Promise<ContentItem[]> {
  if (!screenData.length) return screenData;

  // use ollama's embedding model
  const provider = ollama.embedding("nomic-embed-text");
  const embeddings: number[][] = [];
  const uniqueData: ContentItem[] = [];
  
  for (const item of screenData) {
    // extract text content depending on type
    const textToEmbed =
      "content" in item
        ? typeof item.content === "string"
          ? item.content
          : "text" in item.content
          ? item.content.text
          : JSON.stringify(item.content)
        : "";

    if (!textToEmbed.trim()) {
      uniqueData.push(item);
      continue;
    }

    // generate embedding
    const { embedding } = await embed({
      model: provider,
      value: textToEmbed,
    });

    // check for duplicates using cosine similarity
    let isDuplicate = false;
    for (let i = 0; i < embeddings.length; i++) {
      const similarity = cosineSimilarity(embedding, embeddings[i]);
      if (similarity > 0.95) {
        isDuplicate = true;
        break;
      }
    }

    if (!isDuplicate) {
      embeddings.push(embedding);
      uniqueData.push(item);
    }
  }

  return uniqueData;
}

// cosine similarity helper
function cosineSimilarity(a: number[], b: number[]): number {
  const dotProduct = a.reduce((sum, val, i) => sum + val * b[i], 0);
  const normA = Math.sqrt(a.reduce((sum, val) => sum + val * val, 0));
  const normB = Math.sqrt(b.reduce((sum, val) => sum + val * val, 0));
  return dotProduct / (normA * normB);
}

streaming transcriptions to llms

for real-time analysis of audio:

import { pipe } from "@screenpipe/js";
import { OpenAI } from "openai";
import type { Settings } from "@screenpipe/js";


async function streamMeetingInsights(settings: Settings) {
  const openai = new OpenAI({
    apiKey:
      settings?.aiProviderType === "screenpipe-cloud"
        ? settings?.user?.token
        : settings?.openaiApiKey,
    baseURL: settings?.aiUrl,
    dangerouslyAllowBrowser: true, // for browser environments
  });
  let transcript = "";
  
  // stream audio transcriptions
  for await (const chunk of pipe.streamTranscriptions()) {
    transcript += chunk.choices[0].text + " ";
    
    // analyze every 30 seconds
    if (transcript.length > 200 && transcript.length % 30 === 0) {
      const response = await openai.chat.completions.create({
        model: "gpt-4o", // if user is using screenpipe-cloud, you can also use claude-3-7-sonnet or gemini-2.0-flash-lite
        messages: [
          {
            role: "system",
            content: "provide brief insights on this ongoing meeting."
          },
          {
            role: "user", 
            content: transcript
          }
        ]
      });
      
      console.log("meeting insight:", response.choices[0].message.content);
    }
  }
}

check out our production pipe examples on github to see more ai integration patterns.

notifications (desktop)

await pipe.sendDesktopNotification({
  title: "meeting starting",
  body: "your standup begins in 5 minutes",
})

node.js specific features

the node sdk includes additional features not available in the browser:

// settings management
const settings = await pipe.settings.getAll()
await pipe.settings.update({ aiModel: "gpt-4" })

// inbox management (node only)
const messages = await pipe.inbox.getMessages()
await pipe.inbox.clearMessages()

LLM links

paste these links into your Cursor chat for context:

Getting Started

Reference

Extend

Help

installation

node.js sdk

browser sdk

basic usage

search api

common usage patterns

fetching recent screen activity

searching for specific content

building a timeline of user activity

fetching website-specific content

vercel-like crons

realtime streams

react hooks sdk support ⚛️

using the cli to add hooks

real-time meeting summarizer

real-time screen activity monitor

settings management

quick start with CLI

manual setup

best practices

integrating with llms

vercel ai sdk integration

deduplicating context with embeddings

streaming transcriptions to llms

notifications (desktop)

node.js specific features

LLM links

Getting Started

Reference

Extend

Help

​installation

​node.js sdk

​browser sdk

​basic usage

​search api

​common usage patterns

​fetching recent screen activity

​searching for specific content

​building a timeline of user activity

​fetching website-specific content

​vercel-like crons

​realtime streams

​react hooks sdk support ⚛️

​using the cli to add hooks

​real-time meeting summarizer

​real-time screen activity monitor

​settings management

​quick start with CLI

​manual setup

​best practices

​integrating with llms

​vercel ai sdk integration

​deduplicating context with embeddings

​streaming transcriptions to llms

​notifications (desktop)

​node.js specific features

​LLM links

installation

node.js sdk

browser sdk

basic usage

search api

common usage patterns

fetching recent screen activity

searching for specific content

building a timeline of user activity

fetching website-specific content

vercel-like crons

realtime streams

react hooks sdk support ⚛️

using the cli to add hooks

real-time meeting summarizer

real-time screen activity monitor

settings management

quick start with CLI

manual setup

best practices

integrating with llms

vercel ai sdk integration

deduplicating context with embeddings

streaming transcriptions to llms

notifications (desktop)

node.js specific features

LLM links