Any LLM, locally or cloud

Ollama (free, local), OpenRouter, Gemini, Groq, DeepSeek — or any OpenAI-compatible endpoint. Switch providers with one config line.

12 built-in tools

read_file, write_file, str_replace, patch_file, bash, glob, grep_search, semantic_search, web_fetch, delegate_task, and more — with parallel read-only execution.

Hexagonal architecture

AgentCore is fully decoupled from Node.js. Ports define every I/O boundary — making the core fast to test and easy to port.

Plugin system

Extend with local MCP servers via plugin.json. Add lifecycle hooks, custom slash commands (Skills), and new tools without touching core.

Sub-agents

Delegate subtasks to isolated agent instances. Each sub-agent gets its own context and tool set, up to a configurable depth.

Eval harness

18 LLM-as-judge eval cases covering all built-in tools. Run multi-provider comparative reports to track quality across model upgrades.