A set of 23 community plugins for the hermes-agent framework, built to support an autonomous agent running unattended: self-directed decision-making, observability, cost control, and self-improvement. The plugins install by copying into the agent's plugins directory and share a common utilities module for retry and HTTP logic.
- Autonomy and decision-making: an autonomy engine with decide/plan/reflect tools, a three-model debate "council" for hard decisions, and smart model routing with fallback chains and a sensitivity filter that keeps private data on local models.
- Observability and cost: structured JSON telemetry of every tool call, a unified status tool, an MQTT event stream, and budget enforcement via Langfuse with warnings at spend thresholds.
- Quality and safety: a self-correction loop that critiques outputs with a cheap model, hallucination detection combining regex and LLM scoring, dual-layer prompt-injection screening for email, sandboxed code execution, and session lifecycle management.
- Learning and memory: experiential learning, importance-scored memory with decay, nightly consolidation into MEMORY.md and Qdrant vectors, delegation caching, and a self-updating personality file.
- Communication and integration: a bidirectional bridge to Claude Code (file inbox/outbox plus MCP server), goal management, Telegram UX helpers, an automated research pipeline, daily digests, and delegation quality scoring.
