@maya-ai/tool-agent-runtime

Lean (~280 lines) tool-calling loop driver for LLM agents built directly on top of @mariozechner/pi-ai.

The loop drives multi-turn tool-calling conversations, executes tool batches in parallel by default (or sequentially when any tool declares executionMode: "sequential"), and threads an AbortSignal through every LLM call and every tool's execute method for end-to-end cancellation support.

Install

npm install @maya-ai/tool-agent-runtime @mariozechner/pi-ai

Public API

import {
  runToolAgent,
  type AgentTool,
  type AgentToolResult,
  type AgentMessage,
  type AgentEvent,
  type AgentToolUpdateCallback,
} from "@maya-ai/tool-agent-runtime";

• AgentTool, AgentToolResult — your tool contract: a name, description, JSON-schema parameters, and an async execute(args, ctx) that returns content + optional artefacts.
• AgentMessage, AgentEvent — message and event shapes for streaming agent interactions.
• runToolAgent(input) — runs the agent loop. Accepts a signal?: AbortSignal and emits events through the supplied callback.

Features

• Parallel tool execution — batches run in parallel by default for efficiency
• Sequential mode — tools can declare executionMode: "sequential" when needed
• Full cancellation support — AbortSignal flows through LLM calls and tool execution
• Lightweight — minimal dependencies and focused responsibility

License

MIT.