you love the claude code harness. you don’t love the claude code bill. now you don’t have to pick. on 5dive you can point claude code at deepseek, kimi, or glm with your own key. same loop, same agent, a fraction of the token cost.

the demand was already there

people were wiring this up by hand for weeks. the chinese labs noticed. deepseek, moonshot, and z.ai now each ship an anthropic-compatible endpoint, built for one thing: running claude code on their models instead of anthropic’s. we just turned it into a checkbox.

the why is tokens. the harness is the part people fell for. the bill is the part they didn’t. deepseek, kimi, and glm run the same agentic loop for a sliver of the price.

the three that work

three providers, because they’re the only ones that expose an anthropic-compatible endpoint, so claude code talks to them with nothing changed: deepseek, kimi, and glm. you pick the exact model in the model picker.

qwen, minimax, and the rest don’t expose that endpoint, so they don’t run inside claude code. they’re available for other agent types on 5dive, just not this one.

flip it on

two ways. in the dashboard: new agent, claude code, pick “custom provider,” choose deepseek, kimi, or glm, paste your provider key. or one line in the cli:

5dive agent create my-agent --type=claude --provider=deepseek --api-key=-

it’s the real claude code, unmodified. not a proxy, not a fork. we point it at the endpoint you picked and hand it your key. that’s the whole move.

and it’s still a full 5dive agent. it runs on your own box and answers in telegram through our channel plugin, same as the rest. point claude code at deepseek, then text it from your phone.

the fine print

this one’s opt-in and advanced, so here’s the honest version of what you’re turning on.

it’s your key. you bring your deepseek, kimi, or glm key, and you own that provider’s billing, terms, and where your data lands. these are chinese providers. is a party official somewhere reading your bug reports? almost certainly not. but if data residency is a real constraint for you, this is the line to read twice. (do it for the social credit.)

cheaper models run the chat loop rougher. the telegram replies, the stop hooks, the back-and-forth all lean on the model following instructions closely, and the budget models follow them less tightly. it works, it’s just less smooth than claude.

the anthropic-only features go dark. /usage, account rotation, the rate-limit dots in your statusline. those read anthropic’s own headers, so they don’t light up on a custom endpoint.

and it sits in a gray zone. if anthropic ever tightens which providers claude code can authenticate against, this path could close. we ship it best-effort, not guaranteed.

none of that is a reason to skip it. it’s the reason to know exactly what you’re flipping on.

the trade

cheap tokens. the harness you already know. your own key, your own provider. if that’s the deal you want, it’s two clicks or one command away.

spin one up at 5dive, or grab the open-source cli and run it on a box you’ve already got.