@tpsdev-ai/agent

Native TPS Agent Runtime — headless, mail-driven, sandbox-ready.

The runtime library for building AI agents that run inside TPS branch offices. Provides the core primitives: event loop, mail I/O, memory, LLM provider management, tool registry, and governance boundaries.

Install

npm install @tpsdev-ai/agent

What's Inside

| Module | Description | |--------|-------------| | AgentRuntime | Main lifecycle manager — boot, run, shutdown | | EventLoop | Mail-driven event processing with backpressure | | MailClient | Async Maildir-based messaging (send/receive/queue) | | MemoryStore | Append-only memory with tail-read and size caps | | ContextManager | Workspace context injection and brief generation | | ProviderManager | Multi-provider LLM access (Anthropic, OpenAI, Google, Ollama) | | ToolRegistry | Capability registration with boundary enforcement | | BoundaryManager | Sandbox permission checks — filesystem, network, exec | | ReviewGate | Human-in-the-loop approval for sensitive operations |

Quick Example

import { AgentRuntime, MailClient, MemoryStore } from "@tpsdev-ai/agent";

const runtime = new AgentRuntime({
  agentId: "researcher",
  workspace: "/home/researcher/.tps",
  provider: { model: "claude-sonnet-4-20250514", provider: "anthropic" },
});

await runtime.boot();
await runtime.run(); // starts the mail-driven event loop

Design Principles

• Mail-driven, not chat-driven. Agents communicate through persistent async mail, not ephemeral conversations.
• Sandbox-first. Every operation checks boundaries before executing. Works with nono profiles and Docker isolation.
• No shared memory. Agents are isolated by default. Coordination happens through mail and git, not shared state.

Used By

• @tpsdev-ai/cli — the TPS command-line interface

License

Apache-2.0