Skip to content

UnknownGod2011/DigitalOcean

BranchesTags

Repository files navigation

Cursivis — Cursor-Native AI Agent

Selection = Context · Trigger = Intent · Gradient AI = Intelligence

Built for the DigitalOcean Gradient™ AI Hackathon.

License: MIT

Cursivis turns your cursor into an AI agent. Select text, an image, or a UI region — press a trigger — and DigitalOcean Gradient AI classifies the content, routes it to the right specialist action, and returns the most useful result. Optionally, it executes that result directly in your browser.

What Makes This Different

Most AI tools require you to open a chat window, explain context, paste content, wait, then manually apply the result. Cursivis eliminates all of that.

The interaction is:

  1. Select — highlight text, lasso a screen region, or hold to talk
  2. Trigger — press a button (Logitech MX Creative Console or keyboard shortcut)
  3. Review — result appears in the orb UI and copies to clipboard
  4. Act — optionally press "Take Action" to execute in the browser

The AI decision is visible in every response:

{
  "detectedType": "code",
  "selectedAction": "debug_code",
  "actionLabel": "Debug Code",
  "routingConfidence": "high",
  "routingReasoning": "Content classified as \"code\" → routed to \"debug_code\"."
}

DigitalOcean Gradient AI Features Used

Feature Implementation
Serverless Inference All AI calls via https://inference.do-ai.run/v1/
Claude 4.5 Haiku Text reasoning, intent routing, action planning
GPT-4o Multimodal image + text understanding
GTE Large v1.5 Semantic embeddings and context ranking
App Platform Backend deployment via .do/app.yaml

Single credential: MODEL_ACCESS_KEY from the DigitalOcean Control Panel. No IAM roles, no region config, no proprietary SDK.

Architecture

Logitech MX Trigger / Keyboard Shortcut
        │
Windows Companion App (WPF / .NET 8)
  ├── text selection capture
  ├── lasso screenshot capture
  ├── orb + result UI
  ├── smart / guided modes
  └── voice capture (hold-to-talk)
        │
Gradient Agent Backend (Node.js)
  ├── POST /agent       ← specialist action router + Gradient AI
  ├── POST /analyze     ← legacy companion route
  ├── POST /suggest-actions
  ├── POST /voice       ← voice transcription
  ├── POST /plan        ← browser action plan generation
  ├── POST /embed       ← semantic context ranking
  └── WS   /live        ← real-time voice WebSocket
        │
DigitalOcean Gradient AI (Serverless Inference)
  ├── anthropic-claude-4.5-haiku   (text + routing)
  ├── openai-gpt-4o                (vision + multimodal)
  └── gte-large-v1.5               (embeddings)
        │
Browser Execution Layer
  ├── Chromium extension (current logged-in tab)
  └── Playwright agent (managed browser fallback)

Running Instructions

Prerequisites

  • Node.js 20+ — nodejs.org
  • npm 10+ — included with Node.js
  • DigitalOcean account with Gradient AI Platform access
  • MODEL_ACCESS_KEY — see below

Step 1: Get Your MODEL_ACCESS_KEY

  1. Go to https://cloud.digitalocean.com/agent-platform/serverless-inference
  2. Click the Serverless Inference tab
  3. Scroll to Model Access Keys
  4. Click Create Access Key (or copy an existing one)

Step 2: Configure the Backend

cd backend/gradient-agent
cp .env.example .env

Open .env and set your key:

MODEL_ACCESS_KEY=your_key_here

Step 3: Install Dependencies

cd backend/gradient-agent
npm install

Step 4: Start the Backend

npm start

You should see:

[startup] Validating DigitalOcean Gradient AI Platform connection...
[startup] Endpoint: https://inference.do-ai.run/v1/
[startup] Model   : anthropic-claude-4.5-haiku
[startup] Key     : ***xxxx
[startup] ✓ Gradient AI connection OK — model responded: "ready"
[gradient-agent] Listening on http://127.0.0.1:8080
[gradient-agent] Health: http://127.0.0.1:8080/health

Step 5: Test the Backend

Health check:

curl http://localhost:8080/health

Expected: {"ok":true,"service":"gradient-agent","ts":"..."}

Agent endpoint (text analysis with routing metadata):

curl -X POST http://localhost:8080/agent \
  -H "Content-Type: application/json" \
  -d '{
    "text": "def fibonacci(n):\n    return n if n <= 1 else fibonacci(n-1) + fibonacci(n-2)",
    "mode": "smart"
  }'

Expected response includes:

{
  "detectedType": "code",
  "selectedAction": "explain_code",
  "actionLabel": "Explain Code",
  "routingConfidence": "high",
  "result": "...",
  "model": "anthropic-claude-4.5-haiku"
}

Suggest actions (Guided Mode):

curl -X POST http://localhost:8080/suggest-actions \
  -H "Content-Type: application/json" \
  -d '{
    "protocolVersion": "1.0.0",
    "requestId": "test-1",
    "mode": "guided",
    "selection": { "kind": "text", "text": "Please find attached the Q3 financial report for your review." },
    "context": { "activeApp": "outlook", "cursorX": 0, "cursorY": 0 },
    "timestampUtc": "2026-01-01T00:00:00Z"
  }'

Semantic embeddings:

curl -X POST http://localhost:8080/embed \
  -H "Content-Type: application/json" \
  -d '{
    "query": "machine learning deployment",
    "items": ["Docker for ML models", "Chocolate cake recipe", "Kubernetes for inference", "Neural network training"]
  }'

Step 6: Run the Smoke Test

powershell -ExecutionPolicy Bypass -File .\scripts\smoke-test.ps1 -ModelAccessKey "your_key_here"

Step 7: Launch the Full Demo Stack (Windows)

powershell -ExecutionPolicy Bypass -File .\scripts\run-demo.ps1 -ModelAccessKey "your_key_here"

This starts: backend, browser action agent, extension bridge, and companion app.

Environment Variables

Variable Default Required Description
MODEL_ACCESS_KEY — Yes DigitalOcean Gradient AI Model Access Key
GRADIENT_TEXT_MODEL anthropic-claude-4.5-haiku No Text reasoning model
GRADIENT_VISION_MODEL openai-gpt-4o No Vision / multimodal model
GRADIENT_EMBEDDING_MODEL gte-large-v1.5 No Embedding model
GRADIENT_BASE_URL https://inference.do-ai.run/v1/ No Gradient AI API base URL
PORT 8080 No Backend HTTP port

API Reference

Method Path Description
GET /health Service health check
POST /agent Main agentic endpoint — routing metadata + Gradient AI result
POST /analyze Analyze text/image selection (companion app route)
POST /suggest-actions Get ranked action suggestions for Guided Mode
POST /voice Voice command transcription
POST /plan Generate browser action plan
POST /embed Embed and rank context items by semantic similarity
WS /live Real-time voice WebSocket gateway

/agent Response Schema

{
  "protocolVersion": "1.0.0",
  "resultType": "summary|rewrite|answer|debug|explain|draft|translation|bullets|insights|description",
  "action": "summarize",
  "intent": "summarize",
  "result": "AI-generated output text",
  "alternatives": ["bullet_points", "explain", "translate"],
  "detectedType": "report",
  "selectedAction": "summarize",
  "actionLabel": "Summarize",
  "routingConfidence": "high",
  "routingReasoning": "Content classified as \"report\" → routed to \"summarize\".",
  "latencyMs": 820,
  "model": "anthropic-claude-4.5-haiku",
  "usage": { "inputTokens": 148, "outputTokens": 22 },
  "provider": "DigitalOcean Gradient AI",
  "timestampUtc": "2026-03-18T12:00:00.000Z"
}

DigitalOcean App Platform Deployment

# Install doctl
# https://docs.digitalocean.com/reference/doctl/how-to/install/

# Authenticate
doctl auth init

# Deploy
doctl apps create --spec .do/app.yaml

After deployment, set MODEL_ACCESS_KEY as a secret in the App Platform dashboard: App Settings → Environment Variables → Add Secret → MODEL_ACCESS_KEY

See docs/DEPLOYMENT_DIGITALOCEAN.md for full instructions.

Project Structure

cursivis-gradient/
├── backend/gradient-agent/          # Node.js Gradient AI backend
│   ├── src/
│   │   ├── core/
│   │   │   └── actionRouter.js      # Specialist action router
│   │   ├── services/
│   │   │   ├── gradientClient.js    # OpenAI-compatible DO client
│   │   │   ├── gradientAgent.js     # inferIntent, analyzeSelection, generateActionPlan
│   │   │   ├── gradientVoice.js     # voice transcription + WebSocket gateway
│   │   │   └── gradientEmbeddings.js # embedText, rankOrEmbedContext
│   │   ├── routes/
│   │   │   ├── agent.js             # POST /agent (with routing metadata)
│   │   │   ├── voice.js             # POST /voice
│   │   │   ├── plan.js              # POST /plan
│   │   │   └── embed.js             # POST /embed
│   │   ├── app.js                   # Express app + legacy routes
│   │   ├── server.js                # HTTP server entry point
│   │   ├── gradientService.js       # Factory functions for app.js
│   │   ├── startupCheck.js          # Gradient AI connectivity check
│   │   ├── contentClassifier.js     # Content type classification
│   │   ├── browserActionPlanner.js  # Browser action plan builder
│   │   └── schemas.js               # JSON schema validators
│   ├── .env.example
│   ├── Dockerfile
│   └── package.json
├── desktop/
│   ├── cursivis-companion/          # WPF companion app (.NET 8)
│   ├── browser-action-agent/        # Playwright browser executor
│   ├── browser-extension-chromium/  # Chromium extension (MV3)
│   └── browser-native-host/         # Native messaging bridge
├── plugin/logitech-plugin/          # Logitech MX Creative Console (C#)
├── shared/ipc-protocol/             # JSON schema contracts
├── docs/
│   ├── DEPLOYMENT_DIGITALOCEAN.md
│   ├── HACKATHON_BUILD_POST.md
│   ├── DEMO_SCENARIOS.md
│   └── ARCHITECTURE_DIAGRAM.md
├── scripts/
│   ├── run-demo.ps1
│   ├── smoke-test.ps1
│   └── deploy-do.ps1
├── .do/app.yaml                     # DigitalOcean App Platform spec
├── ARCHITECTURE_PLAN.md
└── LICENSE

Docs

License

MIT — see LICENSE

About

No description, website, or topics provided.

Resources

Readme

License

MIT license
Activity

Stars

0 stars

Watchers

0 watching

Forks

0 forks
Report repository

Releases

No releases published

Packages

 
 
 

Contributors

Languages