Project Overview¶

Simple Explanation¶

Two teammates: - Smart Planner (expensive) figures out the plan - Fast Worker (efficient) executes safe, small steps

ToolWeaver helps the Planner find the right tools and runs many cheap steps in parallel with guardrails. Result: smart planning with fast, low-cost execution.

Technical Explanation¶

Large model plans once (GPT-4o/Claude). Small models and tools (Phi-3/Llama, APIs, sandboxed code) execute many steps. ToolWeaver provides: - Registry & discovery (decorators/templates/YAML + hybrid search) - Parallel dispatcher with limits (cost, time, failures, concurrency) - Sandboxed execution (restricted builtins; timeouts) - Multi-layer caching & idempotency - Observability (metrics/logging)

Architecture at a Glance¶

Natural Language → Large Model (Planning) → Tool Search → Workflow Execution
                   1 call           Narrow K tools      Parallel nodes
                            ↓
     MCP Workers   Function Calls   Sandboxed Code
            ↓            ↓                ↓
     Small Models (Phi-3/Llama) — many cheap calls

Why It Matters¶

Ship faster: remove boilerplate for limits, logging, caching
Scale safely: guardrails prevent runaway cost/fan-out
Stay flexible: decorators, templates, or YAML

Quickstart¶

Install: pip install toolweaver (add extras like [openai], [azure], [anthropic] for LLM providers)
Define a tool and run a parallel demo: see Get Started / Quickstart

Package Extras (what/when/why)¶

azure: Azure AI Vision + Identity — for Azure Computer Vision tools
openai: OpenAI Python SDK — for GPT-4, ChatGPT models
anthropic: Anthropic SDK — for Claude 3, 3.5 models
redis: distributed cache — shared, faster than file cache
vector-db: Qdrant + client — semantic tool search at scale
monitoring: WandB + Prometheus — production observability
all: everything above — one-shot setup

See details: Get Started / Installation