What is context-runtime?

Context Runtime is a Python library that decides what information an AI language model gets to see before it answers a question. The project compares itself to a database query planner, but for AI context instead of SQL: when an application says it needs an answer, the runtime works out what to retrieve, how to compress it, which tools or sources to use, and how to check the result, then records that decision as an inspectable plan it can learn from over time. It is meant for any application that has a decision point about which context or settings to use, along with a way to measure whether the outcome was good. The README lists eleven example use cases, called tenants, each with a small offline test showing the runtime's learned choice scoring higher than a fixed baseline, covering things like picking which skills an assistant should recall, tuning search settings for a retrieval system, or choosing which security data sources to pull for an alert. The project also describes a larger set of fifteen example agent services plus a control layer, each one pairing an existing open source tool, such as a CRM or a monitoring dashboard, with Context Runtime acting as the decision layer inside it. There is also a separate paid enterprise version mentioned that adds extra rules for governance and trust, though it is described as building on top of the open runtime rather than replacing it. Getting started means installing the Python package, with optional extras for connecting real AI models through LiteLLM or for different retrieval methods. The library exposes a small set of core actions in code: run a request end to end, explain how a plan was chosen, or simulate the expected cost and speed without actually running it. There is also a command line interface for running requests from a terminal. The README calls the project a runnable reference implementation and notes it works fully offline using stub models for testing, with real model and retrieval connections available once the right extras are installed.