What Supabase did for databases, neurwerk.base does for AI agents. Twelve integrated layers — frontends, triggers, proxy, LLMs, monitoring, memory, runtime, secrets, security, auth, and inter-agent communication — all fully open-source, self-hosted on your infra.
Built on battle-tested open-source tooling
neurwerk.base is a curated, open-source stack organised into twelve layers. Each layer solves a specific concern — from user-facing interfaces to agent runtime — using the best available open-source tools, pre-configured to work together out of the box.
Central admin UI for managing agents, services, memory, and team access.
User-facing chat interfaces and interaction surfaces for agents.
How agents are invoked: scheduled jobs, webhooks, events, and chat commands.
LLM and MCP routing, load balancing, and HTTP reverse proxy for all agent traffic.
Persistent memory on client, team, user, and agent level. Each level with its own prompt memory, session archives, skills, and external knowledge.
Sandboxed agent execution environment with tool access and lifecycle management.
Custom agentic email messaging system for inter-agent communication.
Encrypted credential storage and secure injection into running agents.
Threat model, network policies, and hardening controls across the stack, with layered PII protection.
User provisioning, SSO, roles, and access rights across the whole stack.
Language model providers available to every agent: local, private, or frontier.
Every layer of the stack comes with production-grade capabilities, pre-integrated and ready to deploy.
SSO, OAuth, JWT, and role-based access for your entire agent fleet, frontends and custom MCP servers. Connect any identity provider. Built on Keycloak.
Real-time token tracking, cost analysis, and usage monitoring per agent and per user. Combined with smart LLM routing across providers, content filters, PII redaction, and jailbreak detection at the proxy layer.
Define, deploy, and connect agents across functions through neurwerk.studio. Use a wide variaty of triggers such as email or text to activate your agents.
Client, team, user, and agent-level memory with prompt injection, full-text session archives, reusable skills, and PostgreSQL/pgvector external storage.
Deploy on AWS, GCP, Azure, or bare metal. No vendor lock-in. Full data sovereignty. EU-first by design.
Ollama, OpenAI, Anthropic, Mistral, Llama or Gemini, one unified interface. Switch models with a config change.
Deploy the full stack today. Self-hosted, open-source, and production-ready.
Star on GitHub →Forever fully open-source.