DISCOVER THE FUTURE OF AI AGENTS

HarnessClaw Engine

Added May 3, 2026
Agent & Tooling
Open Source
Multi-Agent SystemGoAI AgentsWeb ApplicationAgent & ToolingModel & Inference FrameworkDeveloper Tools & CodingProtocol, API & Integration

A Go-based LLM programming assistant backend engine featuring WebSocket multi-turn conversation, tool calling, permission control, skill extension, and multi-agent orchestration.

HarnessClaw Engine is a Go-based LLM programming assistant backend engine that delivers full-cycle programming assistance via WebSocket protocol (v1.4). It can serve as an alternative backend for tools like Claude Code and Cursor, or integrate directly into enterprise workflows through the Feishu channel.

The core engine implements a 5-phase query loop (preprocessing → LLM streaming → error recovery → tool execution → continuation check) with 7 built-in programming tools: Bash, FileRead/Edit/Write, Grep, Glob, and WebFetch. A 6-step permission pipeline (DenyRule → ToolCheckPerm → BypassMode → AlwaysAllowRule → ReadOnlyAutoAllow → ModeDefault) provides fine-grained security control. Context management supports LLM-based auto-compaction with a circuit breaker pattern and configurable token threshold (default 0.8).

The skill system loads definitions from SKILL.md files, supporting YAML frontmatter, parameter substitution, priority overrides, and nested directory discovery. The LLM integration layer uses the Bifrost adapter to unify access to Anthropic, OpenAI, Bedrock, Vertex and other providers, with astron-code-latest as the default model.

Starting from v0.0.5, multi-agent orchestration was introduced, including sub-agent generation (sync/async/fork), a task system, team management, and @-mention routing. The same version added an Artifact storage system with SQLite persistence. v0.0.6 further introduced Agent definition and management (Console HTTP API, skill/tool whitelists, Personality prompt injection).

The architecture follows a strict unidirectional dependency design (Channel → Router → Engine → Provider/Tool), internally decoupled via an in-process event bus, with system prompts assembled from 7 modular sections. The storage layer uses pure-Go SQLite (modernc.org/sqlite) with no CGO dependencies. The project is open-sourced under Apache-2.0, currently at v0.0.6 in early rapid iteration.

Unconfirmed: The iflytek-astron tag suggests a possible connection to iFlytek but is not explicitly stated; Go 1.26+ requirement may be a typo as that version hasn't been released; MCP integration scope (mcp-go in go.mod but undocumented); Feishu channel configuration details not publicly documented.

Related Projects

View All

STAY UPDATED

Get the latest AI tools and trends delivered straight to your inbox. No spam, just intelligence.