Swival

Swival is a coding agent for working with frontier models as well as smaller models, including local ones. It is designed for tight context windows and limited resources, and is described as free, open-source, and easy to set up.

The agent emphasizes context management. It keeps work focused, avoids unnecessary context bloat, and uses graduated compaction together with persistent state so it does not lose track of work under tight limits. It also supports a configurable review loop with LLM-as-a-judge support, and it generates JSON reports that capture timing, tool usage, and context events for comparing models, settings, skills, and MCP servers on real coding tasks.

Swival can auto-discover an LM Studio or llama.cpp model, or be pointed at HuggingFace, OpenRouter, Google Gemini, Vertex AI, ChatGPT Plus/Pro, AWS Bedrock, or any OpenAI-compatible server. It supports interactive back-and-forth sessions, can be embedded through a Python library, and can be exposed as an A2A endpoint over HTTP. The interface also includes commands for learning on the spot and for autonomous security audits that scan a codebase for provable security bugs, then verify findings in isolated agents before reporting them. Extensions are supported through SKILL.md-based skills, MetaSKILLs workflows written in a safe Python subset, and MCP servers.

Installation is shown with uv tool install swival or, on macOS, via Homebrew. The project page also presents a swival --serve mode, an swival --acp mode for ACP-aware editors, and examples of running the agent directly from the command line. The package is presented as a pure Python tool with no framework. It is accompanied by Calibra, described as Swival's companion for benchmarking and evaluation.