Models
Klaus's built-in subscription and custom API model sources, model effort levels for tasks of different complexity, auxiliary task models, and the chat's main model, thinking effort, Codex fast mode, and quota view.
"Settings → Model Sources" connects and manages each source's accounts, endpoints, and model catalogs. "Settings → Model Tiers" configures the high, medium, and low levels and the advisor, while "Settings → Multimodal Models" configures image, voice, music, and video generation. Browsing the settings page does not switch the main model a chat is currently using.
Klaus does not sell models and no membership is required: every source uses your own account or key, and Klaus neither meters usage nor charges licensing fees. Existing sessions that used official membership show the model as unavailable after the upgrade; simply pick another source to continue.
Model sources
There are two kinds of model sources: your own subscriptions, and custom API keys. Each subscription connects differently:
| Source | How to connect |
|---|---|
| Claude Code | Browser sign-in, or reuse this machine's Claude sign-in; several accounts can be saved, see below |
| Codex | Browser authorization (device code), or reuse this machine's Codex sign-in |
| Cursor | Browser sign-in (Cursor must be installed on this machine), or reuse this machine's Cursor sign-in; some models need the matching Max tier |
| Qoder, QwenWork | Browser authorization, or reuse credentials from the official client already signed in on this machine |
| WorkBuddy | Pick the account edition (domestic or international) first, then sign in via browser authorization or reuse the matching sign-in already on this machine; both paths follow the selected edition |
| Antigravity | Sign in with a personal Google account in the browser, or reuse this machine's Antigravity sign-in |
| TRAE CN | Browser sign-in, or reuse this machine's TRAE sign-in; only TRAE CN accounts are accepted, and usage shows as available credits |
| Kiro | Browser sign-in (Google or GitHub); for other sign-in methods, reuse this machine's Kiro sign-in |
| Kimi | Browser authorization, or enter an API key from the console |
| Bailian Token Plan | Enter the plan key and pick the China or international site; you can also sign in to the Bailian console to show Token Plan quota |
| MiniMax M Plan | Enter the plan key; the site can be auto-detected, China, or global |
| Zhipu Coding Plan | Enter the plan key and pick the matching site: China (bigmodel.cn) or international (z.ai); keys do not work across sites |
| Ark Coding Plan | Enter the plan key; you can also sign in to the Ark console to show Coding Plan quota |
Traffic and quota on these sources belong to the corresponding provider account. MiniMax M Plan shares one subscription quota across text, image, speech, and video.
Custom API keys connect your own model service. Providers include Anthropic, OpenAI Compatible, Kimi API, DeepSeek, Google Gemini, AWS Bedrock, and Google Vertex; fill in the API key, endpoint, and model name. A few options that are easy to miss:
- OpenAI Compatible can use the Chat Completions or Responses protocol.
- Anthropic lets you choose which request header carries the key:
x-api-keyorAuthorization: Bearer. Gateways that only accept Bearer work with the latter. - You can declare which reasoning levels the model supports, whether it supports deep thinking, and whether it can read images. For a Qwen-compatible model declared with deep thinking, the chat's thinking switch directly controls whether each answer uses deep thinking.
Multiple Claude accounts and quota rotation
The Claude Code source can hold several accounts. After you are signed in, the "Sign in to Claude" and "Use local Claude login" buttons remain available for adding more accounts; signing in to the same account again only refreshes its credentials.
When the current account is rate-limited or out of quota, Klaus moves it to the back of the queue and retries the same request with the next account, without you switching anything. On an authentication failure, Klaus first tries to renew the sign-in; if authentication still fails and other accounts are available, it tries the next account too. It reports an error after all accounts have been tried. A single account or an entirely expired queue requires a new sign-in. The sidebar's usage entry lists each account's remaining quota and marks the one in use. Usage checks share concurrent requests and cache results briefly. When the usage endpoint is temporarily rate-limited, checks wait before retrying; this does not mean the account's allowance is exhausted.
Model effort levels
The high, medium, and low effort levels configure which model serves tasks of different complexity, and each level can use a model from a different source. For example, the high level can use a Claude subscription, the medium level a custom Qwen, and the low level a small custom model — they do not have to come from the same source as the chat's main model.
- The high level serves the hardest tasks.
- The medium level serves tasks of medium difficulty.
- The low level serves the simplest tasks.
Every new subagent requires an explicit model argument from the delegating main model; users do not need to select it for each delegation. In Settings → Preferences → Configuration, "Choose subagent models from the catalog" is off by default. When off, the main model passes low, medium, or high based on task complexity. When on, it selects a concrete source:model from the shared catalog. This reuses the source and model catalogs, exposing entries from connected sources without guessing model IDs. If the catalog is empty while enabled, new subagents are unavailable with an explanation.
Missing arguments, inherit, and concrete models supplied while the switch is off are rejected. Explicit selections override role and configured child defaults. Invalid tiers, models absent from the catalog, and currently unavailable selections fall back to the main model captured for that creation, with the reason recorded. Saved switch changes apply at the next inference step. Existing subagents keep their original models and continue through followup_task. Context compaction uses its chat's main model. Under "Settings → Model division & orchestration", click "Change" in the "Sub-agent tiers" section, pick a source first, then choose a model and save. When "Follow this chat's main model" is selected, that level uses the main model of the calling chat; when a level is configured but its source or model is currently unavailable, that call falls back to the chat's main model. The "Orchestration" section on the same page chooses the execution model for Orchestrate mode (/orchestrate): the chat model plans and reviews while the chosen model writes code and verifies changes, defaulting to the shared Medium tier with an optional "Fall back to chat model" switch.
Model effort levels and the chat's main model are independent of each other: changing a level does not change the main model selected in the input box. Saving returns immediately while running chats pick up the new configuration in the background. It takes effect on the next call, while requests and sub-agents already in progress keep using the configuration they started with. Models that are signed out, deleted, unauthorized, or that lack a required capability are marked unavailable; new subagents fall back for these selection problems as described above, and record the reason.
Image, video, and speech generation and the advisor are each a single global choice and do not change with the language model effort levels; see Image, Video, and Speech Generation and Advisor. Cross-source calls send the relevant task content to the service you selected. Context compaction processes the session's content.
Auxiliary task models
Chat title generation follows its chat's main model by default. To pin it to another model, pick a source and model under "Settings → Preferences → Conversation"; choosing "Follow the chat model" restores the default. Memory and dream maintenance is configured under "Settings → Dream", message predictions under "Settings → Preferences → Editor", and dictation refinement under "Settings → Voice". All four are global picks that apply to every chat, may cross sources, and are independent of the chat main model and the effort levels.
A task never fails because of its pick: when the configured source is signed out, the model was removed, or the reference is invalid, title generation, predictions, and memory maintenance silently fall back to the chat's main model, and a failed speech refinement keeps the recognizer text. Changes to the first three are pushed to running chats and apply from the next call; speech refinement reads the current pick on each request, so a change applies at the next dictation. Cross-source calls send the relevant task content to the service you selected.
Chat main model and thinking effort
The main model handles direct conversation and tool calls, and is selected separately in the chat input box. The main model's source, model choice, and thinking effort are each saved per chat; changing one chat does not affect other existing chats, and new chats inherit the most recently saved defaults. Switching sources lands on that source's default model (for Codex, GPT-6.1 Sol), which you can then change by hand.
Thinking effort is not the same setting as the high, medium, and low model effort levels. The available levels run from off, low, medium, and high up to extra high and max; which ones appear depends on the capabilities the model declares.
The + menu beside the composer (and the /parallel slash command, with /ultra kept as an alias) provides a standalone Parallel work mode rather than placing it inside the thinking effort menu. While the Parallel pill is active at the bottom of the composer, Klaus splits complex work into parallel sub-tasks and combines the results; requests prefer the model's supported collaboration effort, then Max, then the last declared native value, or Medium when none is declared. Sibling Orchestrate mode (/orchestrate) pairs the chat model as planner and reviewer with an execution model for writing and verifying code. Like the chat model selection, Parallel and Orchestrate modes are saved per chat and remembered as the default for subsequent new chats without changing pre-existing chats (Plan and Goal modes remain scoped to the current chat only); clicking the pill or toggling the mode again leaves it, and leaving Parallel restores your saved native thinking effort. Work modes are independent of permission presets: turning on any mode keeps your current permission selection without extra Full access confirmation dialogs.
You can also change the model, the thinking effort, and the global effort levels while a chat is in progress. Once ready, the new configuration takes effect from the main turn's next reasoning step, without waiting for the final reply. Requests already sent, retries of those requests, and sub-agents, side questions, title generation, and maintenance tasks already started keep using the configurations they captured; switching never ends background preview services or clears chat history.
The input box selector shows the saved choice. Switching models shows no "preparing" prompt — the engine captures the new configuration on the next reasoning step; if preparing or saving fails, the reason is shown inside the input box. When the connection drops or it cannot be confirmed that the engine published the change, Klaus never quietly keeps sending with the old model. Reselecting or retrying checks the engine's record before preparing the latest choice again.
Codex fast mode
When a Codex model is selected, the composer's model menu has a fast mode switch: faster responses at higher usage. The lightning icon before the model name is always shown, outlined when off and filled when on. The setting is saved per chat and also becomes the default for new chats; it has no effect while another source is selected and returns when you switch back to Codex. Other sources do not show the switch.
Message predictions
An empty composer suggests the next message you are likely to send, generated by the model this chat already uses by default, or by a model pinned under Default models → Auxiliary task models — any model source works, with no extra subscription or sign-in. Press Tab to accept it as a draft you can still edit before sending. No prediction appears while a turn is still running or plan mode is on. To turn it off, disable "Show predictions" under "Settings → Preferences → Editor".
Viewing quota
The usage entry at the bottom of the sidebar shows quota for the source the current conversation uses: what remains, when it resets, and the plan; some sources separate personal and organization quota, and sources measured over several windows (such as five hours and weekly) show each window. It follows the conversation you have open and refreshes when you switch; custom API keys show only the model and a billing note.