Why the Stack Matters
The AI agent landscape is evolving at a pace that makes even experienced engineers dizzy. New models, frameworks, and tools launch weekly. Making the right choices — and knowing when to switch — is the difference between a system that ships and one that stalls in perpetual refactoring.
After building dozens of production agent systems, we have developed strong opinions about which tools work in production and which are better left to side projects. This is our current stack and why we chose each component.
LLM Providers
Anthropic (Claude)
Claude is our primary model for most agent tasks. The reasons:
- Instruction following: Claude consistently follows complex system prompts with high fidelity, which is critical when agents need to adhere to specific behavioral constraints
- Tool use reliability: Claude's function calling implementation produces well-structured JSON tool calls with very low error rates
- Long context: Claude's large context window allows agents to process extensive documents and maintain rich conversation histories
- Safety features: Built-in safety guardrails reduce the engineering effort needed for responsible deployment
We use Claude for the majority of our agent work — from customer support agents to document processing pipelines to multi-agent orchestration.
OpenAI (GPT)
We use OpenAI models in specific scenarios:
- Cost optimization: For high-volume, simple classification or extraction tasks where a smaller model suffices
- Fine-tuning: When we need to train a specialized model on domain-specific data
- Client requirements: Some clients have existing OpenAI contracts or preferences
- Multimodal tasks: GPT-4V and similar models for image understanding in agent workflows
Model Selection Strategy
We do not commit to a single provider. Different tasks within the same system may use different models:
- Routing and classification: Smaller, faster, cheaper models
- Complex reasoning: Claude or GPT-4 class models
- Specialized tasks: Fine-tuned models when the volume justifies the investment
- Fallback chains: If the primary model is unavailable, fall back to an alternative
This multi-model approach reduces vendor lock-in and optimizes cost without sacrificing quality.
Agent Frameworks
LangChain
LangChain provides the scaffolding for building complex agent workflows. We use it for:
- Chain composition: Building multi-step workflows where each step can be a different model call, tool invocation, or data transformation
- Tool integration: A rich ecosystem of pre-built tool connectors for databases, APIs, and file systems
- Memory management: Built-in support for conversation memory, summary memory, and knowledge base retrieval
- Evaluation: LangSmith integration for tracking agent performance and debugging failures
MCP (Model Context Protocol)
MCP is our standard for tool integration. Rather than building custom connectors for each tool, we use MCP servers that expose capabilities through a universal protocol:
- Database access: MCP servers for PostgreSQL, MySQL, and MongoDB
- Communication: Slack, email, and messaging MCP servers
- Development tools: GitHub, Linear, and Jira servers
- Custom tools: We build MCP servers for client-specific systems and APIs
MCP's composability means we can add capabilities to any agent just by connecting another server.
OpenClaw
OpenClaw is our go-to for browser-based autonomous agents. When an agent needs to interact with web applications that do not have APIs, OpenClaw provides:
- Browser control: Agents can navigate, click, type, and extract data from any web page
- Local execution: Runs on the client's machine, keeping data private
- Tool ecosystem: Integrates with 50+ tools out of the box
- Extensibility: Custom tool development for specialized interactions
Voice AI
Voice AI is a core differentiator for Snapsonic, drawing on our founder's 20+ years in real-time communications.
LiveKit
LiveKit is our primary real-time communication infrastructure:
- WebRTC native: Low-latency audio and video transport
- AI agent integration: First-class support for AI agents in voice conversations
- Scalable: Handles thousands of concurrent sessions
- Open source: Self-hostable for clients with strict data residency requirements
SignalWire
For telephony integration (making and receiving phone calls), SignalWire provides:
- PSTN connectivity: Connect AI agents to traditional phone numbers
- Programmable voice: Full control over call flow, recording, and routing
- SIP support: Integration with enterprise phone systems
Deepgram
Deepgram handles speech-to-text (STT):
- Low latency: Real-time transcription fast enough for conversational AI
- Accuracy: High accuracy across accents and domains
- Streaming: Continuous transcription as audio streams in, not just batch processing
ElevenLabs
ElevenLabs provides text-to-speech (TTS):
- Natural voices: Some of the most human-sounding AI voices available
- Low latency: Fast enough for real-time conversation
- Voice cloning: Custom voice creation for brand consistency
Frontend & Application Development
Next.js
Next.js is our web application framework of choice:
- React foundation: Component-based architecture with excellent ecosystem
- Server-side rendering: Critical for SEO and performance
- API routes: Backend functionality alongside the frontend
- Vercel deployment: Seamless deployment and scaling
TypeScript
Every line of code we write is TypeScript. For AI applications, type safety is not optional — when your system processes LLM outputs that could be anything, TypeScript's type checking catches malformed data before it causes downstream failures.
Tailwind CSS
Tailwind gives us rapid UI development without design system overhead. For dashboards, admin interfaces, and human-in-the-loop review tools, Tailwind's utility-first approach lets us move fast while maintaining consistency.
Backend & Data
Python
For ML-heavy workloads, data processing, and LangChain-based agent implementations, Python is the right tool:
- ML ecosystem: Unmatched library support for AI/ML workloads
- LangChain: LangChain's Python library is the most mature implementation
- Data processing: Pandas, NumPy, and similar libraries for data transformation
Supabase
Supabase is our default for application databases and authentication:
- PostgreSQL: Full SQL database with pgvector extension for embeddings
- Real-time: WebSocket subscriptions for live updates
- Auth: Built-in authentication and row-level security
- Storage: File storage for documents and media
Vector Storage
For RAG and semantic search, we use pgvector (Supabase's PostgreSQL extension) for moderate-scale applications and dedicated vector databases (Pinecone, Weaviate) for high-volume use cases.
Deployment & Infrastructure
Vercel
Our primary deployment platform for web applications:
- Edge functions: Low-latency serverless compute
- Preview deployments: Every PR gets a preview URL
- Analytics: Built-in performance monitoring
- Scaling: Automatic scaling without configuration
Docker
For services that need more control than serverless — long-running agents, voice AI servers, and background processing — we use Docker containers deployed to the client's infrastructure or cloud provider.
How We Choose Tools
Our tool selection follows four principles:
- Production-proven over cutting-edge: We use tools that have been battle-tested in production, not the newest framework on Hacker News
- Composable over monolithic: We prefer tools that do one thing well and compose with others, rather than all-in-one platforms
- Open standards over proprietary lock-in: MCP over custom integrations, standard APIs over vendor-specific SDKs
- Right-sized for the task: We use expensive, powerful models where quality matters and cheaper models where speed and cost matter more
The stack evolves as the landscape changes. What matters is not the specific tools, but the principles guiding the selection.
Snapsonic builds production AI agent systems using these tools and more. Based in Vancouver, Canada, we help businesses across North America choose the right technology and deploy it effectively. Get in touch to discuss your AI agent architecture.
Frequently Asked Questions
What AI models does Snapsonic use most?
We primarily use Anthropic's Claude models for agent tasks due to their strong instruction following, reliable tool use, and built-in safety features. We also use OpenAI models for cost optimization, fine-tuning, and multimodal tasks. We select models based on the specific requirements of each task.
Why use multiple AI frameworks instead of just one?
Different frameworks excel at different things. LangChain provides excellent orchestration and tool integration. MCP standardizes how agents connect to external services. OpenClaw handles browser automation. Using the right tool for each job produces better results than forcing one framework to do everything.
What is the most important technology choice for AI agents?
The tool integration layer — how your agents connect to external systems and data sources. The LLM model can be swapped relatively easily, but the tool integration architecture determines how flexible and maintainable your system will be. MCP is our recommendation for this layer.
How do you handle vendor lock-in with AI providers?
We design multi-model architectures where different tasks use different providers. This means no single provider outage or pricing change can take down the system. We also use open standards like MCP for tool integration, avoiding proprietary APIs wherever possible.