Multi-Provider LLM Pool
Tag-based routing, automatic failover, provider freeze/thaw, enable/disable toggle. Supports OpenAI, Anthropic, Gemini, DeepSeek, Ollama, Azure, and any OpenAI-compatible API.
Async-first, model-agnostic, fully observable -- built in Rust for production workloads
Tag-based routing, automatic failover, provider freeze/thaw, enable/disable toggle. Supports OpenAI, Anthropic, Gemini, DeepSeek, Ollama, Azure, and any OpenAI-compatible API.
Typed channels, checkpointing at every step, interrupt/resume protocol. Build complex multi-step workflows with conditional branching and parallel execution.
JSON Schema validation, LRU activation, dynamic tool creation. Multi-format parser supporting OpenAI, DeepSeek DSML, MiniMax, GLM4, Longcat, and Qwen3Coder formats.
Short-term, long-term (Qdrant vector store), and working memory with semantic search. Context-aware memory retrieval for rich agent conversations.
Session tree with parent/child delegation. TOML-defined agents with template expansion. Built-in agents for skill ingestion, security, architecture, and more.
Content filtering, PII detection, loop detection, risk scoring middleware. Three-layer defense -- sandbox, middleware interception, human-in-the-loop approval.
8-stage middleware chain for token-budget-aware prompt assembly. System prompt, bootstrap, memory, knowledge, skills, tools, history loading, and context status injection.
Multi-level chunking (L0/L1/L2), hybrid retrieval (BM25 + vector). Import Markdown, code, PDF, and more into searchable collections.
Git-like versioning, experience capture, self-improvement with HITL approval. Skills grow smarter over time through usage patterns.
Web browsing via Chrome DevTools Protocol. Headless or visible mode with full page interaction capabilities.
Expose y-agent as a Discord, Feishu (Lark), or Telegram bot. Platform adapters share the same service container with zero duplication.
Span-based tracing, cost intelligence, trace replay. Know exactly what your agent did, why, and how much it cost.