stagewise

stagewise describes itself as a next-gen agent orchestrator rather than a simple autocomplete plugin. The core idea is "model sovereignty": you choose the model behind every agent, whether that is a frontier model such as Claude, GPT or Gemini, an open-weight model like DeepSeek, Qwen, GLM, Kimi or MiniMax, or something running locally through Ollama, vLLM or llama.cpp. The IDE ships with an optional cloud inference layer so you can start without any API keys, but you can swap in your own keys or a custom endpoint at any time.

The engineering focus is on running agents for long stretches without burning through tokens. stagewise keeps the conversation prefix stable between turns to maximize prompt cache hits and sends a compact state delta when the environment changes instead of rebuilding the whole context. It also compresses context automatically as a task grows, which keeps agents aware of terminal output and browser state at a fraction of the usual cost. The site advertises an average cache hit rate of roughly 87%.

The codebase is published under AGPLv3 on GitHub, the company is stagewise GmbH and is backed by Y Combinator, and an enterprise offering is available for teams that need to run stagewise on their own infrastructure with their own models. The team has also announced Klex Bots, a related product aimed at end-to-end coding tasks from planning through to a merged pull request.