deepnoodle-ai/nvoken

nvoken is an early stage, open source backend service that gives an app the plumbing needed to run AI agents, the kind of assistant that can carry on a conversation, use tools, and complete multi step tasks, without the app's own developers having to build all of that infrastructure themselves. Instead of running its own agent inside your app's code, your app sends nvoken a description of the agent it wants, called a spec, along with the user's message, and nvoken carries out that one conversation turn, handling things like streaming the response back, saving checkpoints so work is not lost if something crashes, and pausing for a human to approve a step when needed. The project's central design choice is that nvoken deliberately does not store the definition of your agents, things like their instructions, their tools, or which AI model they use. Instead, your app is expected to build that description fresh from its own database every time it calls nvoken, so if you support many customers with slightly different agent behavior, you never have to register, update, or migrate a separate copy of each agent inside nvoken itself. nvoken only stores the ongoing conversation sessions, the record of what happened during a turn, and optionally an encrypted copy of the API credentials used to pay for AI calls. The README compares this directly against several other agent hosting options, arguing that most of them either only run on one company's cloud, are not fully open source, or still expect you to register your agent definitions with them, while nvoken avoids all three. Right now nvoken works with Anthropic's and OpenAI's AI models, and the project ships generated client libraries for Go, TypeScript, Python, and Rust so an app in any of those languages can talk to it the same way. It can be self hosted, either as a small packaged deployment on Google Cloud Run with its own database, or as a single combined process you run yourself against your own Postgres database. The project describes itself explicitly as an early implementation, with some pieces, like the full checkpoint and steering system, still being built, and it points readers who find gaps between its stated design and current behavior to open an issue. The code is released under the Apache-2.0 license.