Step‑by‑Step Deployment of Suna: The Open‑Source AI Agent That Beats Manus

What is Suna?

Suna is an open‑source general‑purpose AI agent that supports deep research, sandboxed computer use, browser automation, file management, web crawling, command‑line execution, site deployment, and integration with various APIs.

Technical Architecture

Base model: Claude 3.7 (default), with optional access to OpenRouter and LiteLLM.

Development framework: LangChain and LangGraph.

Sandbox environment: Daytona, which supports MCP tool calls.

Data store: Redis for fast interaction.

Backend services: Supabase for authentication, database, and API layers.

MCP tools: tavily for search and FireCrawl for intelligent web crawling.

Local Installation (Windows)

Create a Conda virtual environment for the backend and install Python 3.11.

conda create -n suna python=3.11  # create Suna backend environment
conda activate suna

Install backend dependencies: pip install -r requirements.txt Install frontend dependencies: npm install Obtain API keys for required services and add them to .env in the backend directory:

Tavily search engine – register on the Tavily website and copy the API key to TAVILY_API_KEY.

FireCrawl – register and copy the API key to FIRECRAWL_API_KEY.

Claude model – obtain an API key and set it in the appropriate variable.

Configure the Daytona sandbox:

Create a Daytona account, build a Suna image, generate a Daytona API key, and add it to .env.

Set up Supabase:

Register on supabase.com, create a project (e.g., suna-test), and note the project reference ID.

Link the local project to Supabase:

supabase login
supabase link --project-ref <PROJECT_ID>

Push the database schema: npx supabase db push Configure model access for the frontend by editing .env.local and adding the DeepSeek API key as OPENAI_API_KEY.

Startup Procedure

Start Redis via Docker Compose:

conda activate suna
docker compose up redis

Launch the backend service:

conda activate suna
python api.py

Run the frontend development server: npm run dev Access the UI at http://localhost:3000 and register with any email to begin interacting with Suna.

Capability Demonstration

A complex task was used to evaluate Suna: retrieve Kaggle competition information from the past six months related to large‑scale multimodal models, extract titles, links, dates, organizers, task types, and resource constraints, and output a structured report (table or JSON). Suna planned the task, executed browser actions, ran command‑line tools, wrote Python scripts, and created files inside the sandbox, demonstrating end‑to‑end autonomous operation.