A wave of local-first, bring-your-own-model coding agents is forming under the cloud harnesses

I keep seeing the same pitch from different repos and it's starting to feel coordinated even though it isn't. nanocoder's tagline is "owe nothing to anyone." AnythingLLM's is "stop renting your intelligence, own it." Kaku and Kilo are circling the same idea. Keep your code and your model on your machine, swap providers whenever you want, never get locked into one vendor's hosted agent. GitHub is where most of these live, and they're all trending at once.

This is the OSS counter-movement to Claude Code, Codex, and Devin. Those are great products. I use Claude Code every day in my personal projects. But the bet these tools are making is that hosted convenience isn't worth surrendering data residency and cost control, and that the gap in raw capability is closing fast enough to make the tradeoff worth it. Heterogeneous local model routing, Ollama, GLM, gpt-oss, all swappable, is shipping as a first-class feature instead of a bolt-on.

The GLM-5.2 story above is what makes this credible. A year ago "run your coding agent on a local model" meant accepting a real quality hit. Now there's a downloadable model beating the export-banned frontier on security work. The capability argument against local-first is eroding in real time. A Hacker News thread this week argued 2026 is the year agents move from demos to infrastructure, and the local-first/local-gateway model is the right architectural bet for privacy and control. Commenters kept naming self-hosted runtimes like Jan and GPT4All.

I'm not fully sold yet, and I want to be honest about why. Running a coding agent well takes more than a good model. It takes the harness, the loop design, the tool integration, the compaction strategy, all the unglamorous orchestration work that Claude Code and Codex have spent a year polishing. nanocoder owning nothing to anyone also means owing nothing to a team that fixes the agent loop for you.

What builders should do: if you have data you legally can't send to a hosted API, or you're tired of a per-token line item that scales with your productivity, set up one of these now and route it at a local GLM or gpt-oss model. Treat it as a real fallback, not a toy. The day a cloud harness has an outage or a price hike or, like Mythos, a government-mandated shutdown, you'll want the local path already wired.