LLM Engineer Toolkit
kalyan-ks‘s “curated list of 120+ LLM libraries category wise” (Apache-2.0, 10.7k★ / 1.7k forks) — 150+ entries across 15 categories, each with a one-line description and a link. It is the whole LLM engineering stack as a directory, and it lands here one day after turingpost-13-frameworks counted the same layer a different way.
The shape of the catalog
| Category | Entries | Category | Entries | |
|---|---|---|---|---|
| LLM Agents | 23 | LLM Data Extraction | 8 | |
| LLM Training and Fine-Tuning | 16 | LLM Inference | 7 | |
| LLM Evaluation | 15 | LLM Structured Outputs | 6 | |
| Others | 15 | LLM Data Generation | 4 | |
| LLM Application Development | 13 | LLM Embedding Models | 3 | |
| LLM RAG | 11 | LLM Serving | 2 | |
| LLM Monitoring | 9 | |||
| LLM Prompts | 9 | |||
| LLM Safety and Security | 9 |
Agents is the largest category, and agents plus application development, prompts and structured outputs is about a third of the catalog — which is why the whole source routes here rather than to the spokes that own its other facets (below).
Two censuses of one layer, and they disagree
turingpost-13-frameworks listed 13 agent frameworks on 2026-08-09. This list names 27 in its Agents section on 2026-08-10. Seven appear on both: CrewAI, LangGraph, Agno, OpenAI’s Agents SDK, smolagents, Pydantic AI, CAMEL.
The disagreement is not random. Six of Turing Post’s thirteen are missing here, and they are almost all of the same kind:
- The vendor SDKs — Google ADK, Microsoft Agent Framework, AWS-defaulted Strands.
- The non-Python ones — Mastra (TypeScript-native), and LiveKit for realtime voice.
- LlamaIndex, which this catalog files under Application Development rather than Agents.
Running the other way, this list carries a whole community stratum Turing Post has none of: AutoGen, Swarms, Swarm, Composio, Browser Use, Atomic Agents, AgentStack, AgentOps, Langroid, Memary, Lagent, BeeAI and more.
So one census is shaped like the vendor landscape and the other like the Python open-source community, and the category “agent framework” has no agreed membership. Neither states an inclusion rule. That is the finding: a reader taking either list as the field would get a different field.
The Anthropic absence, now twice. Yesterday’s page recorded that Anthropic appeared nowhere in
Turing Post’s thirteen, and refused to explain it because no criteria were stated. This catalog has no
Anthropic agent library either — across all 150+ entries, “Anthropic” occurs once, inside the
description of traceAI, a monitoring tool that instruments Anthropic among a dozen other providers.
Two independent lists of different shape, one day apart, and the claude-agent-sdk is in neither.
That is a second data point, not an explanation; both lists are criteria-free and neither is evidence
about the market.
The facets that belong to other spokes
Routed whole and ingested for its agent substance, with the rest recorded here rather than fragmented:
- Training and fine-tuning (16) + data generation and extraction (12) —
../machine-learning-wikiowns train-time and the toolchains that run it. - Inference (7) + serving (2) —
../llm-inference-wikiowns serve-time mechanism. Notably this is the catalog’s smallest pair of categories, which for a directory aimed at “LLM engineers” is a statement about where the work is. - RAG (11) — research-wiki holds RAG as the alternative to fine-tuning.
- Safety and security (9) — jailbreak evaluation, guardrails, red-teaming (Garak, NeMo Guardrails,
LLM Guard, DeepTeam);
../ai-governance-wikiowns the policy side, and the tooling side is thin everywhere in this hub. - Evaluation (15) — the second-largest category after agents, and a quiet rebuke to this spoke’s standing complaint. The corpus keeps finding that nobody evaluates agent tooling; the ecosystem has built at least fifteen libraries for doing it (Ragas, DeepEval, Giskard, Opik, AgentEvals, TruLens…). Tools existing is not evidence they are used, and that gap is worth naming precisely.
Tier
T3. A maintainer-curated directory with per-entry descriptions and a maintained category scheme,
which puts it above a bare link dump. But there is no inclusion criterion, no date on any entry, no
licence or maintenance signal per row, and no indication when the list was last reconciled — so
absence proves nothing and presence proves only that someone added it. Same grade and same reason as
awesome-agent-evolution. freshness: volatile — the star count and the category counts are a
2026-08-10 reading.
Related
agent-framework · turingpost-13-frameworks · awesome-agent-evolution · adk · langchain · claude-agent-sdk · kalyan-ks