HarnessClaw Engine: Go-based LLM Assistant with WebSocket

What changed

HarnessClaw Engine has been released as an open-source project, providing a robust Go-based framework for developing LLM programming assistants. The engine is designed with a core "5-Phase Query Loop" that orchestrates the interaction between the user, the LLM, and tools. This loop includes preprocessing with auto-compaction, LLM streaming calls, error recovery mechanisms like exponential backoff, parallel or serial tool execution, and a continuation check.

The engine supports communication via the WebSocket protocol, utilizing a "UI-first card model" for streaming responses. This model defines actions such as card.add/set/append/tick/close and prompt.user/reply, along with session.event, enabling rich, interactive user experiences. It also facilitates server-side and client-side tool execution, along with permission controls and prompt reviews.

Key features include:

Core Functionality

• 5-Phase Query Loop: Preprocessing, LLM Streaming, Error Recovery, Tool Execution, Continuation Check.
• WebSocket Protocol v2.2 (Card Model): UI-first streaming with 8 actions and 13 card kinds.
• Tool System: Includes 7 built-in tools: Bash, FileRead, FileEdit, FileWrite, Grep, Glob, and WebFetch.
• Permission Pipeline: A 6-step system (DenyRule, ToolCheckPerm, BypassMode, AlwaysAllowRule, ReadOnlyAutoAllow, ModeDefault) supporting 6 permission modes.
• Skill System: Allows loading skills from SKILL.md files with YAML frontmatter, parameter substitution, and priority overrides.
• Multi-Provider Support: Integrates with Anthropic SSE client and the Bifrost adapter for OpenAI, Bedrock, and Vertex AI.
• Context Compaction: LLM-based conversation summarization and a circuit breaker pattern to manage token usage.
• Session Management: Thread-safe session state, multi-connection fan-out, and idle timeout reclamation.

The project is built with Go and includes comprehensive testing capabilities, with unit tests, coverage reports, and integration tests requiring an LLM API. The configuration is managed via a config.yaml file, offering over 50 default settings for server port, WebSocket settings, LLM provider details, and engine parameters.

Why it matters for builders

This release provides builders with a foundational engine to construct sophisticated AI programming assistants. The Go implementation offers performance benefits, while the modular design and extensive feature set, including tool calling and permission management, allow for the creation of complex, secure, and interactive AI agents. Developers can leverage the WebSocket-based card model for building responsive UIs that integrate seamlessly with LLM capabilities.

Practical impact

Developers can now integrate HarnessClaw Engine into their projects to build custom LLM applications. The project's GitHub repository provides quick start guides for building and running the engine, along with instructions for testing. Builders can explore the 7 built-in tools, experiment with the permission pipeline, and extend the engine's functionality through its skill system. The multi-provider support allows flexibility in choosing LLM backends. The project is available under the Apache-2.0 license, encouraging adoption and modification.

Caveats and source limits

The source provides detailed technical specifications and architecture of the HarnessClaw Engine. However, it does not include specific benchmark results comparing its performance against other LLM engines. Information regarding pricing for any associated services or specific LLM models is also absent. The release status is indicated as "fresh release" with version v0.0.18, suggesting it is an early-stage project. The provided GitHub repository is the primary source of information, with no external reviews or independent analyses available.