London‑based AI startup Mindstone has launched Rebel, a local‑first operating system for autonomous agents that automatically matches each subtask to the most appropriate model—whether a cheap cloud service, a powerful reasoning engine, or a locally run model for sensitive data. The platform stores its state, prompts and memory in plain markdown files, giving teams transparent, inspectable instructions and a shared‑memory layer that keeps the right model in the loop without exposing corporate information.
Rebel’s architecture sidesteps the heavy wiring of frameworks like LangGraph or AutoGPT by keeping all configuration in simple .md files, which reduces token overhead and lowers API costs. The system can split a workflow across multiple models, routing routine steps to inexpensive models and reserving high‑cost or local inference for complex or confidential operations. Mindstone distributes Rebel under a “Fair Source” license—free for up to 100 concurrent users, with a paid enterprise tier beyond that, and a two‑year sunset clause that converts each version to MIT‑licensed code.
Early deployments include a 250‑person rollout at Epignosis, where Mindstone reports the equivalent of eight full‑time roles reclaimed over 12 weeks. An Impact Dashboard lets leaders monitor productivity gains without spying on individual work, and Linux support is slated for future release.
