About HexClaw
A local-first, privacy-preserving, full-stack open source AI Agent toolchain ecosystem. Engineered for determinism, making large language models your dependable assistants.
Team
HexClaw is developed and maintained by the hexagon-codes team. We are committed to building a local-first, privacy-preserving, full-stack open source AI Agent toolchain so everyone can use large language models safely and efficiently.
Most existing open source AI agents lack adequate safeguards, rate limiting, and token budget controls, making them risky for production and serious personal workflows. Commercial alternatives provide compliance but rely on closed-source cloud hosting where data privacy is out of your hands.
We combine the functional breadth of open source, enterprise-grade safety, multi-channel connectivity, and 100% local-first deployment into a single cohesive platform.
Open Source Ecosystem
The project uses a layered architecture across 7 repositories with strict one-way dependencies; each layer can be used independently:
-
toolkit ↗ Foundation Layer140,000 lines of pure Go standard library with zero external dependencies; providing caching, crypto, net, and worker pools.
-
ai-core ↗ Capability Layer63,000 lines unified multi-provider adapter layer (OpenAI, Claude, DeepSeek, Ollama), embeddings abstraction, and Smart Router.
-
hexagon ↗ Agent Engine234,000-line core framework with 10 Agent topologies, graph orchestration (Graph/BSP), 7-stage RAG, and <10ms native sandbox.
-
hexclaw ↗ Service Layer752,000 lines single-binary backend; integrates 13 message channels, 6-layer security gateway, and 71 endpoints.
-
hexclaw-desktop ↗ Desktop Client412,000-line cross-platform desktop app built with Tauri; provides chat, workflow canvas, K12 Socratic tutoring, and local knowledge assets.
-
hexclaw-ui ↗ Web UILightweight responsive web client providing multi-device chat, tool visualization, and workflow editing.
-
hexagon-ui ↗ ObservabilityAgent observability console supporting end-to-end tracing, prompt debugging, token cost tracking, and metrics.
Technical Architecture
HexClaw is built on our in-house Hexagon Agent engine with a Tauri + Go stack, engineered around high concurrency, zero bloat, built-in security, and native sandboxing:
5-Layer Strict One-Way Dependencies
From the bottom toolkit, up to ai-core, hexagon, hexclaw, and clients. No cyclic dependencies, allowing every layer to be reused as an independent Go module.
Go Concurrency & Single Binary
Agent workloads are I/O bound. Goroutines handle massive concurrency effortlessly; servers compile to a single self-contained binary running in <100ms with zero Python virtual env headaches.
Native Lightweight Sandbox
2,713 lines of Go leveraging OS primitives (macOS Seatbelt, Linux Landlock/seccomp, Windows AppContainer) for <10ms cold start execution isolation without Docker.
6-Layer Fail-Closed Security Gateway
Auth verification → Rate limiting → Token budget → 10-pattern prompt injection detection → 7-class PII redaction → RBAC and full audit trails.
10 Agent Types & Runnable[I,O]
From BaseAgent to Supervisor teams. Every component implements the unified Runnable[I,O] generic interface, supporting Graph BSP and 7-stage pluggable RAG.
13 Channels & 100% Local-First
Connects natively to Feishu, DingTalk, WeCom, Telegram, Web, and more. Data and vector embeddings stay strictly on your local disk with optional complete offline Ollama operation.
Contact
For business collaboration, open source contributions, or technical inquiries, reach out through our official channels:
- Business Inquiries [email protected] Commercial partnerships and custom integrations
- Support & Feedback [email protected] Bug reports, feature requests, and help
- GitHub Community github.com/hexagon-codes Source code, issues, and pull requests