The open-weight coding agent stack

Three of the four tools have a free tier to start; Claude Code is paid from day one. The 03 COSTS table above breaks down each member's pricing. Three meters climb with use: OpenRouter's routed pay-as-you-go, at each provider's list rate; Kimi Code's weekly credit pool, where heavy use meets the cap; and Claude Code's per-token frontier metering. Those frontier tokens cost most each, so keep that seat for the hard fraction and route routine work through the open-weight lanes: the split, not any single subscription, sets the bill.

Rarely. Kimi Code alone is the daily driver, and a team can run on it for weeks. The numbered steps double as the adoption order: add GLM when you want a second open-weight model to fall back on, a config change not a new tool; bring in OpenRouter once routing models on cost becomes a recurring decision; keep Claude Code in reserve for ambiguous specs and multi-agent work the open-weight lanes stall on.

Your Claude Code subscription already covers the hard fraction: ambiguous specs and multi-agent work, where the closed frontier model still leads. Keep it for that. This stack adds a cheaper daily driver beneath it: Kimi Code and GLM carry routine coding at open-weight rates, and OpenRouter turns model choice into a cost decision, not a fresh integration. The shift is proportion: Claude Code becomes the seat you reach for only when the open-weight lanes stall.

Claude Code and Kimi Code both run as coding agents you drive yourself, the one real overlap here. The dividing rule is how hard the task is. Kimi Code wins the routine, well-specified majority: clear tickets, refactors, tests, at open-weight cost. Claude Code wins the ambiguous fraction, where the spec is thin or the work needs multi-agent coordination. The Claude Code and Kimi Code comparison, linked above, covers the finer calls.

Two signals. First, when the jurisdiction question stops being something you weigh and becomes a hard rule: code and prompts must never leave infrastructure you control, which a hosted API cannot promise. Second, when the open-weight models have become your sustained default and you want the runtime itself under your control, not rented per call. Both point to the world-class tier, the self-hosted coding stack, where you run the weights yourself.

The floor is set twice. Kimi and GLM's hosted APIs are served from China, so regulated or sensitive code should not travel through them; GLM's self-hosted weights keep everything on infrastructure you control. Claude Code needs commercial or API access, not a consumer plan, before proprietary code goes in, since consumer plans train on data unless opted out. OpenRouter adds an intermediary: confirm which providers a route touches before sending regulated data. Review every agent's output before it ships.