At OSS Cafe, we design developer-first, secure, and highly scalable building blocks. From local model access proxy layers to interactive client-side interfaces, we power autonomous systems with confidence.
A unified collection of production-grade libraries, frameworks, and tools designed to secure, monitor, and stream interactive states for modern AI agent integrations.
An enterprise control plane and policy-enforcing reverse proxy for Model Context Protocol (MCP) traffic. Securely manage connection scopes, authenticate client agents, apply Cedar policies to arguments, and monitor metrics in real time.
A portable, framework-agnostic package pairing standard React frontends with .NET backends to stream live interactive components, structured tool outputs, and shared application states using the open AG-UI protocol.
A powerful orchestration engine for multi-agent systems and stateful long-running workflows. Easily implement complex decision trees, human-in-the-loop approval triggers, and event-based execution pipelines.
A secure reverse proxy and control plane for Model Context Protocol (MCP) servers. Regulate tool access, authenticate agents, and audit logs.
Intelligent agents need access to tools to execute tasks, but giving an LLM raw terminal or database access poses critical security risks. Prism MCP Gateway acts as a secure boundary, evaluating every tool request against local rules.
It intercepts standard stdio or SSE messages, decodes the tool arguments, runs them through the fast Cedar policy engine, and blocks unauthorized behavior before it reaches your backend systems.
# Run gateway proxy container
docker run -d -p 8080:8080 \
-v /etc/prism/policies:/app/policies \
ossbrewers/prism-mcp-gateway:latestpermit (
principal == Agent::"Writer",
action == Action::"WriteFile",
resource == Directory::"/var/tmp"
);A stateful orchestration engine purpose-built for long-running, multi-agent systems with human-in-the-loop controls.
Each agent workflow is modelled as a finite state machine. Runs survive crashes, restarts, and deployments — state is persisted and automatically recovered from a durable store.
Workflows can pause mid-execution and surface interactive approval steps to human operators — payment approvals, document reviews, or policy escalations — before continuing automatically.
A high-throughput internal message bus connects workflow steps, spawning parallel sub-agents and coordinating results back into a single coherent execution graph.
Deeply integrated with Prism MCP Gateway for tool gating and Prism Generative UI for surfacing approval widgets — the full stack works as one connected system.
AI-Driven Workflows is planned for release in Q4 2026. Watch the repository for early previews and design documents.
Watch on GitHub →Click any action below to trigger a simulated AG-UI stream. Watch how text outputs, tool calls, and interactive widgets blend into a single, cohesive user experience.
Interact with the simulation by choosing a prompt. The agent uses the AG-UI protocol to stream text and execute custom tool calls, rendering custom interactive UI widgets inside the chat.
We believe that security, transport protocols, and interactive agent interfaces are critical infrastructure. They belong in the public domain.
Every core feature is built fully in the open under permissive MIT licenses. We don't gate basic security, caching, or resilience modules behind enterprise-only pricing models.
Our code structures rely heavily on clean architectural boundaries. Swap out databases, message queues, and client runtimes without rewriting business logic.
We gate release candidates on how fast a developer can boot and run them. If it takes longer than 15 minutes to run an end-to-end example, we don't ship it.