Gen AI Toolkit
Provider-agnostic tool calling and structured output framework
Lightweight, production-grade toolkit for building generative AI applications with unified tool calling, structured outputs, and provider abstraction across OpenAI, Anthropic, Gemini, and local Ollama models.
Visual evidence
System visualization
Case study
Problem, Build, Outcome
Problem
Every LLM provider has divergent tool calling formats, schema definitions, and streaming protocols, forcing teams into vendor lock-in or fragile wrapper code.
Build
Engineered an extensible TypeScript toolkit that normalizes tool execution, schema definition, retries, and output streaming into a single predictable interface.
Outcome
Enables developers to swap models from cloud providers to local instances with zero downstream code changes, with full type safety.
Architecture
System Flow
Architecture Notes
- Adapter pattern decouples application logic from provider-specific wire protocols.
- Schema normalizer converts standard JSON Schema to provider-native tool definitions.
- Execution loop manages multi-step tool calls with recursion limits and error recovery.
Engineering choices
Technical Decisions
Designed around standard Web Fetch and Streams APIs for runtime portability across Node.js, Bun, and edge environments.
Implemented strict tool execution timeouts and cancellation tokens to prevent runaway agent loops.
Separated tool declaration from tool execution for safe sandboxing and permission gating.
Evidence
Proof Points
Unified tool-calling abstraction across 4 provider families
Zero-dependency core runtime with optional provider adapters
Comprehensive test coverage including streaming and timeout failure modes
Next read