Ethos: An Open-Source Agent Framework Built for Local Models
myth007 · reddit · 2026-08-06
A developer released Ethos, an open-source MIT-licensed agent framework, noting that most existing frameworks treat local servers simply as 'OpenAI with a different base URL,' which is the root cause of many local setup failures. Ethos makes five core design decisions specifically for local models:
- Small-context mode: Instead of pasting memory and skills into the prompt, it injects an index and a memoryread tool, allowing the model to pull only what it needs.
- Native structured output: Sends the native format for different backends (e.g., format for Ollama, guidedjson for vLLM).
- Probes real context: Checks the actual served context length at startup rather than blindly trusting advertised numbers.
- Prefix-stable prompts: Places static content at the front and dynamic content at the tail, ensuring the prefix is byte-identical to actually hit prefix caching.
- Sensible timeouts: Retains a 10-minute client deadline to accommodate long periods of no response when local servers load weights into VRAM.
More from coding & agent
- x402 Standard Gains Linux Foundation Backing, AWS & Google Support for AI Agents — 0xJeff · 2026-08-06
- Meta Launches MuseCode Coding Agent Environment with 95% Off Data-Sharing Tier — 量子位 · 2026-08-06
- yacine Aims to Build True RL Game Agents with DeepSeek and Physics — yacineMTB · 2026-08-06
- Reverse Engineering Rain World's Animation System and Building a Custom Renderer Solely with DeepSeek — yacineMTB · 2026-08-06
- Yacine Reverse Engineers Animation System & Builds Renderer Using Only DeepSeek — yacineMTB · 2026-08-06
- Qwen3.8-Max Tested: Top-Tier Frontend Skills, Excels at Composite Agent Tasks — karminski3 · 2026-08-06