Agent Upstart

Cool things people are building with agents — and practical help getting the most out of them.

Playbooks · Coding tasks

Run Kimi, GLM and DeepSeek in the harness you have

Written November 13, 2026

Umans Code serves Kimi K3.1 Pro, GLM 5.3 and DeepSeek V4.1 Flash to Claude Code, OpenCode, Cursor and more — one key, per-token prices.

The setup shape

Umans Code runs three open-weight models — DeepSeek V4.1 Flash, Kimi K3.1 Pro and GLM 5.3 — on its own GPU infrastructure, behind one API key. The tools it names are the ones developers already run: Claude Code, Cursor, Zed, Copilot, OpenCode, Kilo Code, Pi and more. You add credit, create a key, and point the tool you already run at Umans instead of replacing it.

Setup, in the guide's own summary, is one command or two environment variables. The OpenCode path is the lightest: Umans AI Coding Plan is listed on models.dev, so OpenCode finds it natively — add it under Providers and paste the key, or use the CLI route the guide gives. The Claude Code and related paths use the Umans CLI, which opens a browser to authenticate on first run and saves the key. A manual route also exists: the base URL is https://api.code.umans.ai, with an Anthropic-style endpoint and an OpenAI-style endpoint, and umans-coder as the model name to start from. Keys are created in the dashboard under API Keys, and a new key is shown only once.

The model table

All three carry the same 1M context figure in Umans' table, so the choice between them is price and reasoning behaviour, not window size. The spread is wide: DeepSeek's flash model costs a fraction of the Kimi tier per output token, which is why Umans makes it the default rather than the premium pick.

The routing caveat

Two model IDs do not name a fixed model. umans-coder, the recommended default, and umans-flash, the fast tier for supporting roles, route to whatever Umans currently picks. Umans states plainly that the picks change as it evaluates models: today both run DeepSeek V4.1 Flash at its rates, and umans-coder previously ran a GLM flash model and Kimi K2.7-Code before that. The live target is published through the service's model-info endpoint.

That is a reasonable default for a developer who does not want to track model releases, and a real caveat for one who does. A workflow tuned against one model's behaviour can find itself running another model under the same ID. If reproducibility matters, pin one of the three named model IDs instead of the routing IDs.

What to watch

Related playbooks

Sources

Back to: Coding tasks playbooks hub · Also: Run MiniMax in Claude Code