Three repositories near the top of GitHub's trending lists over the past week each pull a different layer out of the coding-agent stack.
- zai-org/ZCode (Apache-2.0, about 7.1k stars and 2.1k forks at time of writing) is Z.ai's coding-agent harness, now open source: an Electron desktop client, a React web client, an HTTP/WebSocket server with remote SSH/WSL support, and a terminal agent CLI/TUI with its runtime. Its latest update, v3.14.3, landed on September 23. A Chinese model lab publishing its full harness -- desktop, web and terminal clients, not just weights -- joins the open-source agent clients already available from OpenAI (Codex) and Google (Gemini CLI).
- yetone/magpie (MIT, about 1.6k stars) is summed up by its tagline: "Codex on DeepSeek, Claude Code on Kimi." It runs a local gateway on
127.0.0.1:3425that speaks the OpenAI chat-completions, Anthropic Messages and Gemini APIs, translates between them (streaming and tool calls included) and forwards each request to whichever vendor serves the chosen model. It lists support for Claude Code, Codex, Gemini CLI, OpenCode, Cursor, Copilot CLI and more. - dzhng/jevgrep (MIT, about 1.5k stars) lets an agent ask natural-language questions about a repository ("how are database connections pooled?") and returns files, reading leads and verbatim excerpts. It installs into Claude Code, Codex or OpenCode as a skill via
jg skill. The project's own benchmark claims about 28-30% lower agent cost on its test tasks.
The trend is decoupling. A year ago, a coding agent's harness, model and retrieval came as one vendor's bundle. These projects treat each as a replaceable part: pick a harness, route it to any model, and bolt on a retrieval skill. That's good for cost and for avoiding lock-in, and it's the open-source side of the vendor skill-pack trend we covered on September 26.
The engineering caveats follow from the same decoupling. Harnesses are tuned for particular models. Running Claude Code's prompts and tool schemas against a different model changes the behaviour you evaluated, and changes which safety training sits underneath your agent's permissions. A local gateway holding every provider key is an attractive target, and magpie sends daily basic telemetry unless you set DO_NOT_TRACK=1. Agents also read their config at startup, so a model switch doesn't reach running sessions. Star counts move daily, and jevgrep's savings figure is self-reported. Treat all three as promising, not proven.
Trending agent tooling is splitting the coding-agent stack into interchangeable harness, model and retrieval layers -- cheaper and less locked-in, but every swap changes the behaviour and safety assumptions you tested, and a key-holding local gateway becomes a target worth defending.