GoClaw is built with production constraints in mind. Explore our 12 foundational systems designed for scale, efficiency, and robust agent behaviors.
GoClaw runs on an always-on loop comprising 8 explicit stages: context → history → prompt → think → act → observe → memory → summarize. Stages are fully pluggable, ensuring predictable run sequences and structured output.
Optimize prompt context windows dynamically. Supports 4 execution modes (Full, Task, Minimal, None) with robust section gating, LLM prompt cache boundary alignments, and dynamic per-session resolution.
Ensures agents retain critical facts. Progressive loading of 3 tiers: Working Memory (immediate chat logs, L0) → Episodic Memory (session summaries, L1) → Semantic Memory (vectorized knowledge graph, L2).
A native document registry supporting wiki-style reference link parsing ([[wikilinks]]), robust filesystem synchronization, and advanced hybrid search utilizing pgvector combined with Full-Text Search (FTS).
Coordinate multiple agents via shared task boards and dynamic delegation pathways (sync, async, and bidirectional). Includes 3 out-of-the-box orchestration modes: Auto, Explicit, and Manual control.
Agents gather performance metrics on tool execution success, analyze their errors, propose configuration modifications, and safely update their own behavior and communication profiles under system-enforced guardrails.
Enforce clean isolation inside a shared database deployment. GoClaw isolates workspaces, segregates user files, maps API credentials securely using AES-256-GCM, and applies precise Role-Based Access Control (RBAC).
Integrate Anthropic natively (custom HTTP+SSE client supporting prompt caching optimizations), alongside OpenAI, OpenRouter, Groq, DeepSeek, Gemini, Mistral, xAI, DashScope, and standard OpenAI-compatible targets.
Bridge agents directly to Telegram, Discord, Slack, Zalo OA, Zalo Personal, Feishu/Lark, and WhatsApp, enabling users to invoke agent routines, schedule cron triggers, and run code files from their messaging feeds.
Hardened out-of-the-box infrastructure featuring a deny-by-default execution engine, prompt injection mitigation, outbound SSRF block gates on HTTP tools, secure key encryption, and per-tenant rate limit enforcement.
No bulky runtimes or complex node_modules trees. The entire GoClaw agent engine compiles down to a single 25MB static executable. Run it with SQLite locally or PostgreSQL on standard VPS environments with <1s cold start.
Track performance bottlenecks. GoClaw features native LLM call span tracing, prompt caching performance metrics, system resource utilization, and optional export pipelines supporting standard OpenTelemetry (OTLP).
