Agent setup index

Each Agent maintains its configuration on its own setup page. That page explains what the Agent does and provides One-time launch, Persistent use, removal, and everyday launch instructions. When a mode is unavailable, the page explains why.

One-click import reads its default metadata and marked command blocks directly from these English setup pages. Console administrators do not copy the commands into a second store: a manual record with the same platform ID is an explicit override, while a new platform ID is a manual addition. Removing an override immediately restores the documentation default.

The {{api_base_url}}, {{access_key}}, {{model}}, and Agent-specific configuration placeholders are publication-template variables. At render time every client configuration is generated from the same effective reasoning capabilities instead of guessing from model names in shell scripts. {{agent_reasoning_manifest_json}} / {{agent_reasoning_manifest_base64}} also expose a versioned, client-neutral capability manifest for future adapters.

Command block markers

Every published command block declares the platform and usage mode as two independent attributes. agent-import-platform is posix for Bash or Zsh on macOS, Linux, and WSL, or windows for native PowerShell; agent-import-command is temporary or persistent. Platform and lifecycle are intentionally not combined into values such as windows-temporary.

Removal commands are documentation-only lifecycle actions, not one-click import templates. Native Windows cleanup blocks therefore use the separate agent-import-action="cleanup" marker together with agent-import-platform="windows"; they never reuse agent-import-command.

In WSL, a POSIX template installs and configures the Agent inside that Linux distribution and its $HOME; it does not modify a native Windows desktop application's settings. Use the Windows tab for a native Windows installation. Agent pages that explicitly target a macOS application remain macOS-only.

The annotation's wslEnabled field makes this distinction machine-readable. It defaults to true for backward-compatible CLI records and is explicitly false for ChatGPT Desktop imports that depend on /Applications, osascript, and open -a. The console then shows those entries as macOS-only instead of advertising WSL support.

For example, a Windows cleanup block is marked agent-import-platform="windows" agent-import-action="cleanup".

Reasoning capability mapping

Capabilities narrow through three layers: native model support comes from models.dev, OpenRouter, or an administrator declaration; transport support comes from the channel type or explicit reasoningTransport; a routed downstream model then takes the intersection of every candidate. An Agent adapter may narrow this result again, but can never invent a tier that the model or route cannot carry.

Transport declarations are structured and traceable. An administrator may override the mapping and carried efforts, but the override is intersected with native model support and never silently substitutes a lower effort. Codex imports also send the versioned X-Lumoswitch-Agent: codex-v1 marker so the LiteLLM compatibility layer does not have to rely only on request-shape detection.

Omission is a controlled outcome, not missing handling. If native support is unknown, the channel cannot transport the control, client vocabulary does not intersect, or no stable client field exists, the template sends no reasoning parameter and avoids The inference request is invalid.

Choose an Agent

Return to Agent setup for editor Agents. SDKs and ordinary graphical clients are listed under API and compatibility in the sidebar.