Install Footprint Guide
wasmagent packages are composable — install only what you need.
Package footprint table
| Package | Install size | Edge (Cloudflare Workers) | Notes |
|---|---|---|---|
@wasmagent/core | lightweight | ✅ | No native deps; tree-shake aggressively |
@wasmagent/mcp-firewall | lightweight | ✅ | Pure TypeScript; no native deps |
@wasmagent/aep | lightweight | ✅ | Pure TypeScript; Zod only |
@wasmagent/mcp-policy | lightweight | ✅ | Thin wrapper over mcp-firewall |
@wasmagent/mcp-attestation | lightweight | ✅ | Pure TypeScript |
@wasmagent/mcp-gateway | lightweight | ✅ | Pure TypeScript |
@wasmagent/compliance | lightweight | ✅ | Pure TypeScript |
@wasmagent/capability-compiler | lightweight | ✅ | Pure TypeScript |
@wasmagent/otel-exporter | lightweight | ✅ | HTTP only; no native deps |
@wasmagent/evals-runner | lightweight | ✅ | Pure TypeScript |
@wasmagent/kernel-quickjs | medium (~1 MB) | ✅ | WASM binary included |
@wasmagent/kernel-wasmtime | medium | ❌ Node.js only | Requires Node WebAssembly API |
@wasmagent/kernel-pyodide | heavy (~8 MB+) | ❌ | CPython-in-WASM; too large for Workers |
@wasmagent/kernel-remote | lightweight | ✅ | HTTP only; connects to E2B/Cloudflare |
@wasmagent/tools-browser | heavy | ❌ Node.js only | Requires Playwright |
@wasmagent/tools-rag | medium | ⚠️ partial | Vector DB connectors vary by backend |
@wasmagent/tools-web | lightweight | ✅ | HTTP only |
@wasmagent/model-local | heavy | ❌ Node.js only | node-llama-cpp native bindings |
@wasmagent/model-anthropic | lightweight | ✅ | Thin SDK wrapper |
@wasmagent/model-openai | lightweight | ✅ | Thin SDK wrapper |
@wasmagent/model-deepseek | lightweight | ✅ | Thin SDK wrapper |
@wasmagent/model-doubao | lightweight | ✅ | Thin SDK wrapper |
@wasmagent/model-qwen | lightweight | ✅ | Thin SDK wrapper |
@wasmagent/model-zhipu | lightweight | ✅ | Thin SDK wrapper |
@wasmagent/model-minimax | lightweight | ✅ | Thin SDK wrapper |
@wasmagent/model-moonshot | lightweight | ✅ | Thin SDK wrapper |
@wasmagent/aisdk | lightweight | ✅ | ai (Vercel AI SDK) is a peer dep |
@wasmagent/mastra-sandbox | lightweight | ✅ | @mastra/core is a peer dep |
@wasmagent/openai-agents | lightweight | ✅ | @openai/agents is a peer dep |
@wasmagent/claude-agent-sdk | lightweight | ✅ | @anthropic-ai/claude-agent-sdk is a peer dep |
@wasmagent/react | lightweight | ✅ | react is a peer dep |
Edge compatibility detail
Packages marked ✅ work in Cloudflare Workers with no special configuration. Packages marked ❌ require a Node.js environment (local, server, or CI). @wasmagent/kernel-remote is the recommended isolation layer for edge/serverless deployments.
Optional peer dependencies
Install these only when you use the corresponding adapter:
| Package | Optional peer | When you need it |
|---|---|---|
@wasmagent/aisdk | ai | Using the Vercel AI SDK integration |
@wasmagent/openai-agents | @openai/agents | Using the OpenAI Agents JS SDK |
@wasmagent/claude-agent-sdk | @anthropic-ai/claude-agent-sdk | Using the Anthropic Claude Agent SDK |
@wasmagent/mastra-sandbox | @mastra/core | Using Mastra as the agent framework |
@wasmagent/tools-browser | playwright | Browser automation tools |
@wasmagent/model-local | node-llama-cpp | Running local LLMs without outbound traffic |
Quickstart install footprints
Trust Pack (minimal — evidence layer only)
npm add @wasmagent/mcp-firewall @wasmagent/aep~50 KB installed. Covers: tool protection, evidence emission, CI gate.
Runtime (with sandbox execution)
npm add @wasmagent/core @wasmagent/kernel-quickjs~1.3 MB installed (includes WASM binary). Covers: sandboxed code execution, tool calling agents.
Full stack (runtime + evidence + framework adapter)
npm add @wasmagent/core @wasmagent/kernel-quickjs @wasmagent/mcp-firewall @wasmagent/aep
npm add @wasmagent/aisdk # or openai-agents / claude-agent-sdk / mastra-sandbox~1.5 MB + framework peer deps. Covers: sandboxed execution, evidence, firewall, framework integration.
Compliance research
npm add @wasmagent/compliance @wasmagent/aep @wasmagent/evals-runner
pip install evomergeCovers: constraint verification, evidence export, training data generation.
Model SDK install cost
All model adapters are thin wrappers. The actual SDK is installed once as a direct dep:
npm add @anthropic-ai/sdk # for @wasmagent/model-anthropic
npm add openai # for @wasmagent/model-openai
# etc.@wasmagent/model-local with node-llama-cpp downloads model weights separately (not part of npm install). Model weights are typically 1–8 GB and are cached in ~/.cache/node-llama-cpp.