Base Layer EP 03: Spencer Kimball on Governance Is the Real Blocker

Listen now

Build with permission-aware AI tools.

Tools, plugins, and reference implementations that bring SpiceDB's fine-grained permissions to coding agents, IDEs, MCP servers, and RAG pipelines. All open source.

SpiceBox

Runtime governance for AI coding agents. Wraps Claude Code with three enforcement layers — deterministic permission checks, OS sandboxing, and a network filtering proxy — so coding agents are constrained to the right tools and resources.

• Hook server, OS sandbox, and network proxy. Defense-in-depth around the agent runtime.
• Every tool call resolves against a real SpiceDB permission check — not a YAML allowlist.
• One command to wrap a sandboxed session: spicebox claude.

authzed/spicebox Public

Fine-grained permissions for AI coding agents

Read the launch post

⠀⠀⠀⠀⠀⢀⣀⣀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⣠⣤⣶⣤⡀
⠀⠀⠀⠀⢠⣿⣿⣿⣿⣦⡀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠠⢀⣾⣿⣿⣿⣿⡇
⠀⠀⠀⠀⠸⣿⣿⣿⣿⣿⣿⣌⢢⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⡰⢡⣿⣿⣿⣿⣿⡿⠃
⢀⣤⣴⣶⣶⣬⣛⠻⢿⣿⣿⣿⡆⢳⡰⣶⠶⢂⣀⡰⠶⠦⣰⠁⣿⣿⡿⠟⣋⣁
⣾⣿⢩⣴⣶⣶⣶⣶⣶⣦⣝⠻⣿⠈⣷⣶⣿⣿⣿⣿⣿⣷⣯⡘⠟⣩⣴⣮⢻⣿⡆
⣿⣿⢸⣿⣿⣿⣿⣿⣿⣿⣿⣷⡌⡜⠉⠉⢻⣿⣿⣿⡟⠁⠀⠑⠸⣿⣿⣿⢸⣿⡇
⣿⣿⢸⣿⣿⣿⡟⠉⠉⢻⣿⣿⡇⡇⠀⠀⢀⣿⣿⣿⣧⠀⠀⢀⡇⣿⣿⣿⢸⣿⡇
⣿⣿⢸⣿⣿⣿⣿⡄⢠⣾⣿⣿⣧⡻⣦⣤⣾⣿⣄⣼⣿⣷⣶⠟⣱⣿⣿⣿⢸⣿⡇
⣿⣿⢸⣿⣿⣿⣿⣧⣼⣿⣿⣿⣿⣿⣶⢍⣙⣿⣿⣿⣿⣩⡕⣾⣿⣿⣿⣿⢸⣿⡇
⣿⣿⠸⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⡇⣬⣍⢿⣿⣿⣿⢋⣭⡘⣿⣿⣿⡿⢸⣿⡇
⠘⠿⣷⣶⣶⣶⣶⣶⣶⣶⣶⣶⣶⣶⣶⣬⣥⣶⣶⣶⣶⣦⣭⣶⣶⣶⣶⡿⠟

$spicebox claude

spicedb-dev

The official Claude Code plugin for SpiceDB. Plan, implement, and audit SpiceDB schemas from your editor. Includes custom agents to integrate SpiceDB into your app.

• Design, test, and refactor SpiceDB schemas with AI guidance.
• Generate app code to integrate SpiceDB and add permission checks.
• When using Claude Code, add the AuthZed Marketplace, then /plugin install spicedb-dev@authzed-marketplace.

authzed/authzed-marketplace Public

Official Claude Code plugin marketplace for AuthZed

definition document {
  relation editor: user
  permission view = editor
}

MCP servers

Reference implementations of remote MCP servers with spec-compliant authorization built in. Pair with the AuthZed and SpiceDB Dev MCP servers — see the docs for installation in Claude Desktop, Cursor, and other MCP-aware clients.

• Reference implementation with Next.js, BetterAuth, and SpiceDB. Starting point for production-ready MCP servers.
• MCP OAuth implementation and examples of scoped tokens and permission checks at tool-call boundaries.
• Use the AuthZed MCP servers to explore SpiceDB and develop against local SpiceDB instances.

authzed/mcp-server-reference Public

Reference implementations of MCP (Model Context Protocol) servers that include spec-compliant authorization and permission handling for secure AI tool access.

langchain-spicedb

Permission-checked retrieval for RAG pipelines. A drop-in retriever wrapper that enforces SpiceDB checks on docs. Works with LangChain, LangGraph, and any vector store.

• Apply permission checks against retrieved docs. Ensure only documents the user is authorized to access are processed.
• Vector-store agnostic: Pinecone, Chroma, Weaviate, pgvector — wherever your embeddings live.
• One install, drop-in replacement for your existing retriever: pip install langchain-spicedb[all].

authzed/langchain-spicedb Public

Authorization library for RAG (Retrieval-Augmented Generation) pipelines using SpiceDB. Designed for LangChain and LangGraph integrations with support for any vector store.

Retrieved documents → SpiceDB filter → authorized context

Browse the org

Every project above lives in the AuthZed GitHub org alongside SpiceDB itself.