DevCopilot
edge-native AI assistant for Cloudflare Workers development
Project metadata
Overview
An AI development assistant for Cloudflare Workers running entirely on Cloudflare's edge infrastructure. Llama 3.3 70B via Workers AI, stateful conversation memory in Durable Objects backed by SQLite, and Zod-validated tool calling for error analysis, code review, and documentation search.
Architecture
Components
Technical details
Model integration
Uses @cf/meta/llama-3.3-70b-instruct-fp8-fast via Workers AI:
The model runs on Cloudflare's GPU infrastructure with FP8 quantization.
Durable Object state management
Each user session gets a dedicated Durable Object instance with SQLite storage:
Tool calling architecture
Tools are defined with Zod schemas for type-safe validation:
Documentation search is pattern matching plus KV-cached lookup across 30+ topics, not vector retrieval.
Project structure
Bindings (auto-bound via wrangler.toml): AI (Workers AI), CACHE (KV namespace), COPILOT_AGENT (Durable Object).
Requires Node.js 22.12+ (or 20.x), a Cloudflare account with a Workers paid plan for AI access, and Wrangler CLI 4.x+.
Full stack
Cloudflare Workers · Workers AI · Durable Objects · SQLite · KV · React 19 · Vite · Tailwind CSS 4 · TypeScript · Zod · Llama 3.3 70B · Wrangler
Results
Deployed and live | Yes | |
Model config | Llama 3.3 70B FP8, max_tokens 2048, temperature 0.7 | |
Per-session Durable Object with SQLite | Yes | |
Error patterns handled | 12+ | |
Documentation topics | 30+ | |
Zod-validated tool schemas | 3 tools |