Open Source Radar — November 20: Magpie runs Claude Code on Kimi

Today's open-source signal is in the seams around the agent, not the agent itself: which model it runs on, who investigates the build it just broke, and whether the harness wrapped around it can improve without touching the weights.
Magpie (Go, 39,365 stars, MIT) is yetone's local gateway for a mess every agent user now lives in: Claude Code keeps its model in one file, Codex in another, each with its own key and base URL, and a subscription you pay for works in exactly one of them. Magpie runs a small server on your own machine that speaks OpenAI, Anthropic and Gemini in both directions and translates between them — streaming, tool calls and reasoning included — then rewrites the single config key that matters in each of more than 20 agents, so Claude Code on Kimi is a menu-bar pick rather than a hand edit. When a quota runs out it moves on to the next account or model instead of surfacing the error, and an optional masking layer swaps keys, passwords and emails for placeholders before a request leaves the machine. Eight weeks old and already carrying 4,764 open issues: that number is interest, not polish, and the README is upfront that you don't need it if one agent on one vendor's plan is enough.
OpenSRE (Python, 13,320 stars, Apache 2.0) is a CLI agent that treats a red build as an investigation rather than a button to press again. When CI fails it pulls the failing checks and logs, recent commits and deploys, and correlated metrics and traces, then tests hypotheses in a tool-calling loop and separates failures CI caused from failures the code caused; it pushes a fix to the pull request branch for review and posts the summary to Slack, PagerDuty or Telegram. More than 60 integrations cover observability, infrastructure, databases and incident tooling, it runs self-hosted against whichever model you point it at, and masking of sensitive identifiers is available before any external call. The project ships a release every day — today's is dated November 20 — and labels itself a public alpha, which reads as honest: the investigation and the evidence trail are the working parts, the automated fixing is the one to watch. Worth a look if your team's answer to a flaky pipeline is still to run it again.
Harness² (Python, 24,528 stars, Apache 2.0) is Google Research's framework for the question that has quietly become the real one in agent work: not which model, but which harness — the prompts, tools, memory and orchestration wrapped around it. It refines that wrapper with the weights frozen, at two levels: within a task it proposes harness edits, probes them on the same task, and composes the improved version from the difference in behaviour, with no ground-truth answers to lean on; across tasks a second harness keeps what worked so the next task starts better. The reported gains are modest but real — a single step lifted GLM-5 and Gemini 3.5 Flash by up to 11.5 points on WorkBuddy-Bench, 4.5 on JobBench and 11.8 percentage points on LAB's partial pass rate, per the paper. It ships with OpenCode and Codex executors and three professional benchmarks, which is the difference between a research result and something you can point at your own eval.
Worth watching this week: whether magpie's failover holds as vendor terms tighten, and whether Harness²'s gains survive outside its own three benchmarks.
If your agent's model choice is still a config file you edit by hand, you are the audience for all three. Which seam in your stack still needs a tool? Write to us via our contact page.







