Skip to content
thndrs

Umans Provider

Models

Section titled “Models”

The Umans integration includes:

  • umans-coder - the recommended route, currently backed by Kimi K2.7-Code.
  • umans-kimi-k2.7 - hard coding tasks with always-on reasoning.
  • umans-glm-5.2 - GLM 5.2 and the largest Umans context window.
  • umans-flash - a fast model for short context, summaries, and interactive work.

Umans can change the model behind umans-coder; use model metadata or /model to inspect the current catalogue.

Authentication

Section titled “Authentication”

Create an API key in the Umans dashboard under API Keys. thndrs does not accept API keys in CLI flags. You can either export UMANS_API_KEY for a process-local credential or run the guided login flow:

Terminal window
thndrs setup --provider umans


# or use a process-local credential:
export UMANS_API_KEY=sk-...


# or, in an interactive terminal:
thndrs login umans

The guided flow asks whether the key belongs in the global store (~/.thndrs/credentials.env) or the current project store (.thndrs/credentials.env). Project credentials are added to Git’s local exclude file. Keys are never written to TOML, sessions, prompt inspection, snapshots, or diagnostics.

Messages API

Section titled “Messages API”

The provider client uses the Anthropic-compatible Umans Messages API:

  • Base URL: https://api.code.umans.ai
  • Endpoint: POST /v1/messages
  • API key header: x-api-key
  • Version header: anthropic-version: 2023-06-01

Streaming Events

Section titled “Streaming Events”

Provider streaming output is normalized into app events. Assistant text, reasoning, tool calls, completion, and errors are kept as distinct event types so the transcript can render them separately.

Extended thinking is a provider/model capability. For models that support the toggle, thndrs lowers On and None to the Messages API thinking object; Auto leaves Umans’ model default unchanged. Models with always-on reasoning only expose Auto in the picker.

Tool Schemas

Section titled “Tool Schemas”

Local read-only tools are sent as provider-native tool schemas. Tool descriptions stay short and focus on purpose, safety limits, and truncation behavior.

Error Mapping

Section titled “Error Mapping”

Provider errors are mapped into actionable transcript errors and return the prompt draft to a usable state.

  • Authentication failures point to UMANS_API_KEY/thndrs login umans. If UMANS_API_KEY is an environment override, replace or unset it before login.
    • A stored key cannot take precedence over it.
  • rate limits suggest checking Umans usage
  • network and server failures can be retried