The Atomic Agent team explained how their desktop app (v0.0.1) works purely as a client of the agent's OpenAI-compatible HTTP server rather than a private UI pipe. The post documents the endpoints, streaming extensions, mid-turn steering, and approval side-channel that any third-party client can reuse. It is a developer preview, so API behavior may still change between releases.
The open-source MIT terminal/desktop agent Atomic Agent added a hard rule in its Fusion mode: the planner model can read, delegate, and reply, but cannot mutate anything during its turn, because earlier softer gates kept letting the planner do worker work itself. The post explains the failures that led to the rule and ships importers for six other agents.
The open-source MIT terminal/desktop agent Atomic Agent added a hard rule in its Fusion mode: the planner model can read, delegate, and reply, but cannot mutate anything during its turn, because earlier softer gates kept letting the planner do worker work itself. The post explains the failures that led to the rule and ships importers for six other agents.
The makers of Atomic Agent, an MIT-licensed open-source agent, ran one autonomous build — five physics scenes on one page — through three setups to test the popular 'cheap local model + expensive cloud model' pattern. Sonnet 5.5 alone scored 4/5 scenes for $1.83 in 9 minutes; splitting planning (Sonnet 5.5) from coding (local Qwen 3.8 27B on an RTX 3090) cut cost 2.7x to $0.67 but dropped quality to 2/5 and stretched runtime to 85 minutes. The savings came almost entirely from output tokens, since the planner writes briefs instead of code, and all setups passed headless console checks despite visibly broken scenes.
The makers of Atomic Agent, an MIT-licensed open-source agent, ran one autonomous build — five physics scenes on one page — through three setups to test the popular 'cheap local model + expensive cloud model' pattern. Sonnet 5.5 alone scored 4/5 scenes for $1.83 in 9 minutes; splitting planning (Sonnet 5.5) from coding (local Qwen 3.8 27B on an RTX 3090) cut cost 2.7x to $0.67 but dropped quality to 2/5 and stretched runtime to 85 minutes. The savings came almost entirely from output tokens, since the planner writes briefs instead of code, and all setups passed headless console checks despite visibly broken scenes.
An Android app that runs Gemma 4 E4B entirely on-device on a Pixel 10 Pro XL and exposes it as an OpenAI-compatible HTTP server. Any OpenAI client (e.g. TypingMind) can use it via /v1/chat/completions with streaming. It solves the need for private, offline LLM inference without cloud or subscription for Pixel owners.
An Android app that runs Gemma 4 E4B entirely on-device on a Pixel 10 Pro XL and exposes it as an OpenAI-compatible HTTP server. Any OpenAI client (e.g. TypingMind) can use it via /v1/chat/completions with streaming. It solves the need for private, offline LLM inference without cloud or subscription for Pixel owners.
At AI Engineer World's Fair 2026, OpenAI Codex DX engineer Dominik Kundel presented the Codex App Server, an open-source JSON-RPC protocol wrapping the Codex harness that powers the Codex app, IDE extensions, and third-party clients. Developers can build custom agents on it with ChatGPT sign-in, custom instructions, dynamic tools, and a Python SDK.
At AI Engineer World's Fair 2026, OpenAI Codex DX engineer Dominik Kundel presented the Codex App Server, an open-source JSON-RPC protocol wrapping the Codex harness that powers the Codex app, IDE extensions, and third-party clients. Developers can build custom agents on it with ChatGPT sign-in, custom instructions, dynamic tools, and a Python SDK.
An orchestration layer on top of llama.cpp's ggml-rpc backend that lets mixed home machines (CPU-only laptop, CUDA PC, Metal Mac) share memory to serve a 13B LLM. It handles node discovery, memory-aware sharding, and failover, exposing an OpenAI-compatible API. Aimed at people who want to run larger models locally without a single big GPU.
An orchestration layer on top of llama.cpp's ggml-rpc backend that lets mixed home machines (CPU-only laptop, CUDA PC, Metal Mac) share memory to serve a 13B LLM. It handles node discovery, memory-aware sharding, and failover, exposing an OpenAI-compatible API. Aimed at people who want to run larger models locally without a single big GPU.