OpenClaw Classes Map to Different AI Models

What It Is

OpenClaw takes an unconventional approach to model selection by mapping AI capabilities to character classes. Instead of choosing between model endpoints or API parameters, developers select a game-style character class that determines which underlying model handles their requests.

The Hunter Alpha class connects to MiMo V2 Pro, a text-only model with a 1,048,576-token context window and 32,768-token output capacity. The Healer Alpha class uses MiMo V2 Omni, which processes both text and images but reduces the context window to 262,144 tokens while maintaining the same output limit. The implementation details are documented at https://github.com/openclaw/openclaw/pull/49214.

This design abstracts technical specifications behind thematic choices. Rather than comparing token limits and modality support, users pick a class based on their workflow needs. The Hunter excels at processing massive text corpora, while the Healer handles visual content alongside text.

Why It Matters

This approach addresses a common friction point in AI tooling: model selection paralysis. Developers often face lengthy comparison matrices weighing context windows, pricing tiers, and feature sets. By framing these decisions as character classes, OpenClaw reduces cognitive load during the selection process.

The 4x context difference between classes reflects genuine architectural tradeoffs. Text-only models allocate their entire parameter budget to language understanding, enabling deeper analysis of extensive codebases, legal documents, or research papers. Multimodal models split resources between vision and language processing, necessarily constraining context capacity.

Teams working with monorepos or large documentation sets benefit from Hunter’s million-token window. A typical enterprise codebase might span hundreds of files - having the entire context available eliminates the need for chunking strategies or retrieval-augmented generation pipelines. Healer serves different scenarios: analyzing UI screenshots, processing technical diagrams, or extracting information from scanned documents where visual context matters more than extreme text depth.

The game mechanics metaphor also creates intuitive mental models. Developers familiar with RPG archetypes immediately grasp that Hunters track and analyze (deep text processing) while Healers diagnose and repair (multimodal problem-solving with visual inspection).

Getting Started

OpenClaw requires selecting a class when initializing the client. Here’s a basic implementation:

# For deep code analysis hunter = ClawClient(character_class="hunter_alpha")
response = hunter.analyze(codebase_path="./src")

# For multimodal tasks healer = ClawClient(character_class="healer_alpha")
response = healer.diagnose(
 text="Debug this rendering issue",
 image="screenshot.png"
)

The class choice persists for the session, though developers can instantiate multiple clients for different task types. When processing a large repository, Hunter’s extended context means feeding entire file trees without pagination. For Healer, the workflow centers on pairing textual queries with visual inputs.

Documentation at https://github.com/openclaw/openclaw covers installation and authentication. The project uses standard API key configuration, with class selection happening at the client level rather than per-request.

Context

Traditional AI platforms expose models through version numbers and capability flags. OpenAI offers gpt-4-vision-preview versus gpt-4-turbo, Anthropic provides claude-3-opus with different modality support. These naming schemes prioritize technical accuracy over usability.

OpenClaw’s gamification trades precision for approachability. The downside: developers must learn the mapping between classes and underlying models. Someone familiar with MiMo V2 specifications needs to remember that “Hunter” means the Pro variant. This indirection adds a translation layer absent in conventional APIs.

The context window disparity also creates hard constraints. A task requiring both image analysis and 500K tokens of context cannot use either class effectively. Healer lacks sufficient context, while Hunter cannot process images. Conventional APIs let developers make granular tradeoffs; OpenClaw’s class system enforces binary choices.

Alternative approaches include LangChain’s model-agnostic interfaces or LiteLLM’s unified API, which abstract provider differences without thematic metaphors. These tools prioritize interoperability over conceptual frameworks, appealing to teams that prefer explicit configuration over abstraction layers.