Codex CLI with a local model: LM Studio and Ollama setup, models, troubleshooting

Prerequisites

• Codex CLI: npm install -g @openai/codex (codex --version to confirm).
• A current LM Studio (0.3.29 added the Responses endpoint) or a current Ollama.
• A model with tool calling and a large context window; see the next section for memory. Codex also wants a project directory to work in, and it reads an AGENTS.md there if present, which is the right place for durable instructions.

Choosing a local model

Codex drives the model through tool calls (read, edit, run) and sends a long system prompt, so two properties decide whether a model is usable: reliable tool calling and a context of at least 64K tokens (Ollama's recommendation for Codex, while LM Studio says "more than ~25k", which is the floor rather than a comfortable value). The models below have both, and the memory figures are what the server needs with a 64K context loaded.

openai/gpt-oss-20b (Ollama: gpt-oss:20b)

14 GB download, about 16 GB of memory, 128K context, tool calling and reasoning built in. The default for codex --oss and the sensible first choice on a 24–32 GB Mac or a 24 GB GPU.

qwen/qwen3-coder-30b (Ollama: qwen3-coder)

19 GB, a 30B mixture-of-experts model with 3B active parameters, 256K context. In my use the stronger coder of the two for edits across several files, and it needs 32 GB or more.

openai/gpt-oss-120b (Ollama: gpt-oss:120b)

65 GB, meant for an 80 GB GPU or a 96–128 GB Mac. Noticeably better judgement on larger tasks, but out of reach on ordinary hardware.

google/gemma-4-12b (Ollama: gemma4)

About 7 GB on disk, 12 GB loaded with a 32K context. The option for a 16 GB machine, and it completed file-editing tasks in my tests. Smaller models (ibm/granite-4-micro and similar) run too, but tend to narrate edits instead of making them, as the troubleshooting section describes.

Measured on a 24 GB Mac with a 32K context: gpt-oss:20b loads fully on the GPU, while qwen3-coder fits only partly (16 of 19 GB) and slows down, and a dense 27B model such as qwen3.8 runs out of GPU memory under Codex's prompt. On 24 GB that leaves gpt-oss-20b as the working choice.

With LM Studio

lms load openai/gpt-oss-20b --context-length 65536
lms server start --port 1234
codex --oss --local-provider lmstudio
codex --oss --local-provider lmstudio -m qwen/qwen3-coder-30b   # any other loaded model

Codex talks to LM Studio on the URL http://localhost:1234/v1/responses. Load the model yourself with --context-length as above: if Codex's first request triggers the load (LM Studio does load on demand, and a short task then succeeds), the model gets the server's default context, which is usually far too small, and the session fails with a context error a few turns in. lms ps shows what is loaded and with which context. Download models through LM Studio's catalog (lms get openai/gpt-oss-20b): a GGUF from an arbitrary repository can carry a chat template that fails on Codex's tool definitions ("Error rendering prompt with jinja template").

With Ollama

OLLAMA_CONTEXT_LENGTH=65536 ollama serve     # one terminal (or the app's settings)
ollama pull gpt-oss:20b
codex --oss --local-provider ollama -m gpt-oss:20b   # another terminal

Ollama's default context depends on GPU memory and is 4K on most laptops, which is not enough for Codex, and the context-length section of the Ollama guide for Claude Code explains the setting for the app, the terminal and the Linux service. ollama ps must show 65536 in the CONTEXT column.

Two things to know about the built-in ollama provider. It always talks to localhost:11434 and ignores OLLAMA_HOST, so for a model server on another machine, define your own provider (below). And if the model is not present, Codex pulls it without asking — codex --oss on a machine with an empty Ollama starts a 13 GB download of gpt-oss:20b.

ollama launch codex                 # picks a model, writes a profile, starts Codex
codex --profile ollama-launch       # reuse that profile later

Make it permanent with a profile

Instead of flags, put the choice in ~/.codex/config.toml. The simplest form sets the local provider for --oss:

oss_provider = "lmstudio"        # or "ollama"
model = "openai/gpt-oss-20b"
model_context_window = 65536     # what you loaded; lets Codex compact in time

With that, codex --oss needs no further flags (without it, the interactive CLI asks which provider to use, and codex exec exits with an error). The provider IDs lmstudio, ollama and openai are built in and cannot be redefined. A custom provider is needed only for a server that is not on localhost — a GPU box in the next room, say:

[model_providers.ollama-gpu]
name = "Ollama on the GPU box"
base_url = "http://gpubox.local:11434/v1"
wire_api = "responses"

[profiles.gpu]
model_provider = "ollama-gpu"
model = "gpt-oss:20b"
model_context_window = 65536

Then codex --profile gpu. wire_api = "responses" is the default and the only value Codex accepts: a server that implements only Chat Completions (older llama.cpp, vLLM or mlx_lm setups) answers Codex's requests with 404 and cannot be used directly, while current LM Studio and Ollama serve /v1/responses. Model names must be the server's API identifiers (ollama ls, or LM Studio's Developer tab), not display names. Command-line flags override the file for a single run.

On Linux

Everything above works unchanged on Linux, and only the place where the model runs differs:

• Ollama uses NVIDIA (CUDA) and AMD (ROCm) GPUs and falls back to the CPU when the model does not fit, and ollama ps shows the split (100% GPU is what you want). A 20B model with a 64K context needs a 24 GB card, while 16 GB cards run it partly on the CPU and become slow. The service reads OLLAMA_CONTEXT_LENGTH from systemctl edit ollama, not from your shell.
• LM Studio has a Linux build with CUDA and Vulkan backends, and the same lms commands apply. LM Studio's LM Link can also run the model on a remote machine while Codex runs on your laptop, which is the practical answer when the GPU box is elsewhere.

Troubleshooting

404 on every request

The server does not implement /v1/responses (Chat Completions only, or an old version). Update LM Studio or Ollama, or point the profile at a server that does.

A download starts that you did not ask for

The built-in Ollama provider pulls a model that is not present locally. Stop it with Ctrl-C, ollama pull the model you want, and name it with -m.

stream disconnected before completion: Error rendering prompt with jinja template

The model file's chat template fails on Codex's tool definitions (LM Studio). Use the catalog build of the model (lms get …) or override the template in LM Studio's model settings, because nothing on the Codex side is wrong.

stream disconnected … Insufficient Memory (Metal) or a CUDA out-of-memory error

The model plus KV cache does not fit the GPU. Use a smaller model or a shorter context, which on 24 GB means gpt-oss:20b.

The model is reported as missing

The name is a display label, not the API identifier. Copy it from ollama ls or LM Studio's Developer view (qwen/qwen3-coder-30b, not "Qwen3 Coder 30B").

A context or token-limit error after a few turns

The server runs with its default context. Load the model with --context-length 65536 (LM Studio) or start Ollama with OLLAMA_CONTEXT_LENGTH=65536, then verify with lms ps or ollama ps.

Codex says it edited a file but nothing changed, or prints tool calls as text

The model's tool calling is unreliable. Use gpt-oss:20b or qwen3-coder, keep tasks to one to three files, and ask for the change first and for the test run afterwards rather than for both in one prompt.

Codex forgets a constraint you gave three turns ago

Local models hold instructions less firmly than the hosted ones. Put durable rules in AGENTS.md in the project, and restate the critical constraint at the end of a long prompt.

Slow on Linux

The model spilled to the CPU. Use a smaller quantisation, a smaller model or a shorter context, and ollama ps tells you which of them you need.

Connection refused

No server on the port in the profile. Run lms server status or ollama ps, then check that the base URL in the profile matches.

What to expect

The 20B models handle scoped work — a new function with tests, a refactor within a file, a script, documentation — and lose the thread in large multi-file changes. Give Codex a small, verifiable task, let it run the tests, and read the diff. Compared with Claude Code on the same local model, Codex is more demanding on context and a little more fragile with weak tool callers, but with gpt-oss:20b and 64K of context both work well for the kind of scripting a research group does every day.

Sources: LM Studio, Codex integration · Ollama, Codex CLI integration · Codex configuration reference · Ollama library, gpt-oss and qwen3-coder · Codex CLI local mode with Ollama or LM Studio (ynaito.dev).