Blog
Short technical pieces about model-agnostic agents, open-source LLM control, the ModelRein protocol design, and what you can do when every open-weight model is one HTTP call away.

ModelRein vs MCP: What's the Difference?
Both have "Model" and "Control" in the name, but they solve completely different problems. MCP connects a chat model to external tools; ModelRein reaches into the model itself. Here's the split, with a worked example of using both together.
Agents built against a single model's API break the moment the model changes. We argue that a control protocol between the agent and the model layer is the missing abstraction, and that it's especially needed in the open-weight world where new state-of-the-art models drop every few weeks.
Write the agent once. Point ModelRein at Qwen/Qwen2.5-1.5B-Instruct, meta-llama/Llama-3.1-8B-Instruct, or deepseek-ai/DeepSeek-V3. Same client code, same primitives, three completely different model architectures.
Zero per-model configuration isn't a slogan — it's a loader. Here's what happens inside the server when you point it at a new repo_id: config introspection, layer discovery via separability scan, auto-strength calibration against the residual-stream norm.