LLLM

LLLM, short for Low-Level Language Models, is a protocol and service layer for reusable agentic tactics. Its core idea is a Tactic: a typed unit of work that can run in process, stream results, sit behind a FastAPI service, and be described as a PsiHub package resource. The stated aim is to give different runtimes and surrounding tools a stable boundary for agentic work.

The system centers on a small contract rather than on model execution itself. LLLM is not a model runtime; execution details such as tools, provider settings, tracing, eval hooks, and workflow state are handled by Pydantic AI, native LLLM objects, plain Python, or future adapters. The documentation says LLLM owns Tactic, TacticInfo, CallContext, and TacticEvent, along with local, async, streaming, proxy, and sandbox wrappers at the tactic boundary. It also provides FastAPI service adapters and remote tactic clients, exposing /run, /stream, and /info with stable envelopes.

LLLM also includes metadata helpers that export tactics to PsiHub package resources. Those resources are described as discoverable, configurable, and composable, and the page shows ref resolution for local or remote tactic bindings. The text also notes that package storage, validation, cards, agent cards, config templates, semantic channels, event logs, artifacts, snapshots, local stores, service launch decisions, and operational orchestration stay outside LLLM’s responsibility.

The tool is presented for apps, workers, robots, coding agents, package tools, and similar callers that need to understand the same tactic contract without caring whether the underlying implementation is an offline test agent, a Pydantic AI agent with tools, a native prompt/dialog workflow, or a remote HTTP service. No pricing or license information is stated in the provided material.