Software stack

Component map (what you need)

Local coding is five layers. Missing any required layer = “nothing works.” Optional RAG only if you need repo retrieval beyond the model context.

1. 1

   1. IDE / Agent client

   VS-Code-/Cursor-Erweiterungen (Continue, Cline), JetBrains-Plugins (Continue und ähnliche) sowie CLIs wie Aider. Alle auf einen lokalen OpenAI-kompatiblen Endpunkt.

   Der Wirt ist IDE oder Terminal — das Plugin ist nicht das Modell. Repos auf Agent / Toolchain.

   ↓

2. 2 Optional

   2. RAG / repo index

   Die meisten Clients bringen einen Index mit (Continue/Cline-Codebase, Aider-Repo-Map). Eigene Schicht: AnythingLLM, RAGFlow oder LlamaIndex; Stores Chroma, Qdrant oder FAISS; Embeddings lokal.

   Weglassen, wenn kurzer Complete reicht; bei kleinem Modell und großem Monorepo fast Pflicht. Das ist nicht die Inferenz-Runtime.

   ↓

3. 3

   3. Local runtime (OpenAI-compatible server)

   Ollama, llama.cpp server, MLX, vLLM, or SGLang—exposes localhost chat/completions.

   This is the “stack” decision on this page. Community stars below measure popularity of this layer.

   ↓

4. 4

   4. Model weights on disk

   GGUF / MLX / safetensors from the Models allowlist—size locks your hardware band.

   API-only MoE coders never land here.

   ↓

5. 5

   5. Hardware (VRAM / unified memory)

   NVIDIA CUDA, Apple unified + Metal/MLX, or limited AMD/Intel paths—see Hardware.

   Weights fit ≠ context fit (KV tax).

Quick “what am I missing?”

• Have an IDE but no localhost runtime → install Ollama / llama.cpp / MLX first.
• Runtime up but empty models → download an allowlisted coder GGUF (Models page).
• Model loads but agent is blind in a big repo → add RAG / repo index (optional layer).
• OOM or tiny context → Hardware band is wrong, or KV budget is too high.

Decision tree

Community signal — runtime popularity (GitHub OSS)

Stars from Trends segments inference / serving / llm_lib. This is popularity of layer 3 (local runtime)—not SWE skill, not IDE quality, not model quality.

Use it to sanity-check “is this runtime still alive?”, not to pick a coding model.