Openrouter for models?

Not a dumb idea at all — and there's a supported path for it that doesn't require structured SOPs as a workaround.

The cleanest route today is the Hermes Local adapter, which has multi-provider routing built in and explicitly supports OpenRouter as a provider. From the Hermes Local docs (/docs/adapters/hermes-local), provider accepts auto, openrouter, nous, openai-codex, zai, kimi-coding, minimax, and a few others. You set model in provider/model format, point your OpenRouter key in via a secret ref, and you're routing through OpenRouter on the next heartbeat. As a bonus, Hermes also gives you persistent memory and a 30+ tool suite, which tends to make open-source models punch above their weight on Paperclip workloads.

If you'd rather stay on a CLI you already know, the OpenCode Local adapter also takes provider/model ids and is friendly to alternative providers — see OpenCode Local (/docs/adapters/opencode-local). For anything more bespoke (a self-hosted model, a custom proxy in front of OpenRouter), the HTTP (/docs/adapters/http) and Process (/docs/adapters/process) adapters cover it, and the External Adapters (/docs/adapters/external-adapters) doc walks through packaging your own.

A few things worth thinking about before you do this for cost reasons:

Match adapter to role, not the whole company. You probably don't want every agent on an open-source model — strategic reasoning at the CEO layer is where capability matters most. The pattern most cost-conscious operators land on is capable model at the top, cheaper model for execution. There's a worked example with cost reasoning here: Building a Mixed-Adapter Team (/guides/adapters/mixed-adapter-teams) and a more focused version in Claude vs Codex as a Worker (/guides/adapters/claude-vs-codex-as-worker). The structured SOPs idea is right, but it's not a workaround — it's how mixed-adapter teams stay coherent. When the CEO writes the task and a cheaper worker model executes it, the CEO's task descriptions and any output-format skills are what keep the worker from going off the rails. The "Using skills to normalise output" section of the mixed-adapter guide covers this directly. Before switching anyone, check where your usage actually goes. Open the Costs page, filter by agent, and look at cost-per-run. Often one or two agents account for most of the bill — and the fix is sometimes a tighter instruction file or a shorter context, not a model swap. If the usage limits you're hitting are rate limits rather than spend limits, OpenRouter does help there immediately.

Short version: Hermes Local + OpenRouter is the supported path; mix adapters by role rather than swapping the whole team; and keep the CEO on something capable.