Kalyan KS's LLM Engineer Toolkit is a single README that sorts open-source LLM libraries by the job they do, one row per library: name, a one-line description (often the project's own tagline) and a link. Denis Bakanov's LinkedIn post (2026-11-09) is a tour of it. Treat it as a map of the stack's layers and a shortlist per layer; it has no stars, benchmarks, maintenance status or licence column, so it cannot rank anything.
What it is
The README advertises "120+" libraries; the table rows actually number 177 across 15 sections (counted 2026-11-15; a few names repeat across sections). Application development is split into sub-layers: frameworks, data preparation, multi-API access, routers, memory, interface, low code and cache.
| Section | Rows | Examples |
|---|---|---|
| Application development | 27 | LangChain, LlamaIndex, Haystack, LiteLLM, mem0 |
| Agents | 27 | LangGraph, CrewAI, Pydantic AI, OpenAI Agents SDK, smolagents, Browser Use |
| Training and fine-tuning | 18 | Unsloth, TRL, Axolotl, torchtune |
| Evaluation | 18 | Ragas, DeepEval, OpenAI Evals |
| Others | 13 | mergekit, Vanna, pandas-ai, MarkLLM |
| RAG | 12 | Chonkie, rerankers, sqlite-vec |
| Monitoring | 11 | Phoenix, Helicone, MLflow |
| Safety and security | 10 | NeMo Guardrails, Guardrails, LLM Guard, Garak |
| Prompts | 9 | DSPy, LLMLingua, Promptify |
| Inference | 9 | vLLM, llama.cpp, Ollama, TensorRT-LLM |
| Data extraction | 8 | Crawl4AI, Docling, LlamaParse |
| Structured outputs | 6 | Instructor, Outlines, XGrammar |
| Data generation | 4 | DataDreamer, Promptwright, EasyInstruct |
| Embedding models | 3 | Sentence-Transformers, Model2Vec, Text Embeddings Inference |
| Serving | 2 | LangCorn, LitServe |
The repo is also a funnel: its header promotes the author's interview-questions book (with a discount code) and newsletter. That does not make the list wrong, but it explains the breadth-over-judgment shape.
How to use it
Use it to see what layers exist and to name two or three candidates per layer; then choose with your own criteria. Bakanov says as much: the notes are short, there are no stars or benchmarks per library, "so the picking is still your job."
His sharper point is that the right pick is sometimes no framework. His teams used LangGraph in two products, but wrote a sales agent's pipeline directly on the OpenAI SDK so the model could not skip or loop a step. That is the workflow-over-agent argument (R015, R079): when the steps are known, put control flow in code and let the model fill slots, instead of handing the loop to a framework that lets the model choose.
A practical selection pass per layer:
Decide whether the layer needs a library at all (a 50-line loop, a retry wrapper or a JSON schema check is often enough).
Shortlist from the list, then check each candidate's last release, open-issue trend, licence and who maintains it.
Prefer libraries that sit below your control flow (inference, structured outputs, tracing) over ones that own it (agent frameworks), because the latter are the hardest to remove.
Run one real task from your product through the top two before adopting either.
Gaps and staleness
The list is organised around the 2024 LLM-app stack (framework, RAG, eval, monitoring) and is thin on the layers that agent systems added since. As of 2026-11-15 it has none of these:
Tool protocols and gateways. No MCP servers, MCP gateways or auth (R062, R071), and no A2A or agent-UI protocols (R049). The one MCP mention is a framework tagline.
Sandboxes. No E2B, Daytona, Modal or interpreter-style sandboxes for running agent code (R030, R061).
Coding-agent harnesses. No Claude Code, OpenCode or Aider, though that is where most agent engineering now happens (R023, R054).
Durable execution. No Temporal, Restate or similar for long-running agents (R035).
Common tools. No Langfuse or Promptfoo; OpenTelemetry appears once, in a description.
Some rows are dated. OpenAI's Swarm is listed beside the Agents SDK that replaced it, and AutoGen appears without noting Microsoft's move to a successor framework. Without a last-release column, a reader cannot tell a maintained project from an abandoned one; check each repo before adopting it.
Review checklist
When a design doc says "we'll use library X from the toolkit list":
Does this layer need a library, or is the logic small enough to own?
Does the library own control flow? If so, can the model skip, repeat or reorder a step you need fixed?
When was its last release, and who maintains it? Is it superseded?
Licence, telemetry defaults and transitive dependencies checked?
Was it tried on one real task against an alternative, or chosen from a one-line tagline?
How hard is it to remove in six months (does it leak types and abstractions through your code)?
Related entries
R022 and R023: learning paths for AI engineers; this list is the library index that goes beside them.
R015 and R079: workflows over agents; Bakanov's plain-SDK sales agent is a worked example.
R006 and R054: harness building blocks and how real coding agents are built, the layer the list misses.
R073: another curated GitHub list (security repos for AI-written code); same caveats about taglines over evidence.
R041: constrained decoding, the research behind the structured-output libraries listed.
Sources
Denis Bakanov, "120+ open-source LLM libraries in one GitHub list", LinkedIn, 2026-11-09.
Kalyan KS, LLM Engineer Toolkit, GitHub README.