Frank Denis, creator of libsodium, dnscrypt-proxy, and Pure-FTPd, has released Swival, a local coding agent tool designed for small models running on a user's machine. Built because existing agentic tools "crash" and raise privacy concerns, Swival includes an `--encrypt-secrets` option that detects credentials in messages and encrypts them before they leave the machine. The tool addresses context limitations of local LLMs, which often have "32K of context," by capping tool outputs at the source: "50 KB max per file read, 100 grep results, 100 entries per listing." A snapshot system strips out "12,000 tokens of reads" and replaces them with a "200-token summary." It also features automatic compaction in "7 progressive levels" and work notes that survive all compaction. Swival supports "11 backends," including LM Studio by default (zero config), llama.cpp, HuggingFace, OpenRouter, Google Gemini, Vertex AI, ChatGPT Plus, AWS Bedrock, Apple Foundation Models in experimental mode, a generic OpenAI-compatible provider, and an external command. Installation is a one-liner via uv with Python 3.13 minimum: `uv tool install swival`.