Civic Labs
Civic Labs is where we build and open-source the infrastructure that powers forward deployed agents. Every flask here solves a real problem we encountered deploying production agents for mid-market businesses — identity, authorization, guardrails, threat detection, and knowledge retrieval.
Each project is presented as a flask, representing different stages of development. Learn more about flask status.
Flasks are standalone services, libraries, or tools that can be used independently or composed together. They are built by the same team that has shipped identity infrastructure for the German Federal Chamber of Notaries, enterprise AI for BMW and Panasonic, and production systems across financial services, manufacturing, and publishing.
Explore the flasks below or get in touch if you want to contribute, try things early, or shape what comes next.
Getting Started
Want to try out these experiments? Get started here.
Feedback & Contribution
We're building in the open and welcome community input. Learn how to contribute.
Flasks
Hosted MCP manager that unifies and orchestrates multiple MCP servers with built-in authentication and governance — the core of every Civic agent deployment.
Wrap any MCP server in a configurable security layer. Block, redact, and transform tool calls and responses before they reach your agent.
LLM-based threat detection for prompts and tool calls. Catches prompt injection, data exfiltration attempts, and policy violations in real time.
Middleware hook system for MCP servers. Intercept, inspect, and modify tool calls without changing the underlying server — powers guardrails and custom logic.
RAG-powered AI assistant for internal operations. Query company documentation, processes, and institutional knowledge through natural language.

Concepts & Architecture
Understanding the building blocks behind our experiments.
What is MCP and why it matters for AI agent deployment
How guardrails protect agents and the systems they interact with
Understanding prompt injection attacks and LLM safety
OAuth2, granular permissions, and consent for agent access
A middleware layer around MCP APIs for custom logic
Retrieval strategies for grounding LLMs in your data