Codex CLI (--oss) with a local model: setup for Ollama or LM Studio

Codex CLI is OpenAI's open-source (Apache-2.0) terminal coding agent. Its --oss mode runs against a local Ollama or LM Studio server, and Ollama can launch it with a ready-made profile; it suits people who want OpenAI's agent loop on a machine that can hold a 64k context.

Source codeOfficial docsCompare all agents

At a glance

Type
cli
Licence
Apache-2.0
Latest release
0.163.0 (2026-10-15)
Status
Maintained
Local backends
ollama, lmstudio, openai-compatible
Needs native tool calling
Yes
MCP support
Not verified
Recommended minimum context
64,000 tokens
Start-up prompt
~25,450 tokens (single third-party measurement)
GitHub stars
Not verified

How to set it up with a local model

Pick your backend. Each command has its own Copy button. Where a step uses an example model, swap in the one you run; the table below lists the ones that fit your machine.

  1. Install it (macOS / Linux):

    curl -fsSL https://chatgpt.com/codex/install.sh | sh

  2. Start Ollama with a 64k context window (or set the context slider in the Ollama app's settings).

    OLLAMA_CONTEXT_LENGTH=64000 ollama serve

  3. In a second terminal, pull a tool-capable coding model (replace qwen3-coder:30b with any model you prefer).

    ollama pull qwen3-coder:30b

  4. Quickest path: let Ollama launch Codex with a dedicated profile; /model inside Codex switches between your installed tool-capable models.

    ollama launch codex

  5. Or run Codex in OSS mode against Ollama with a specific model.

    codex --oss -m qwen3-coder:30b

  6. Optionally make Ollama the default local provider for --oss.

    ~/.codex/config.toml
    oss_provider = "ollama"
Other ways to install

npm

npm install -g @openai/codex

Homebrew

brew install --cask codex

Check that it works

Start a session and ask Codex to list the project files. With Ollama, run `ollama ps` in another terminal and check that CONTEXT shows at least 64000 and PROCESSOR shows the GPU.

ollama ps

Which local models fit, by machine

Checked by this site's engine at the context Codex CLI needs, with 32 GB of system RAM for MoE offload. A Mac counts three quarters of unified memory as usable.

MachineModelQuantContextFitOllama
12 GB GPUQwen 3.5 9BQ4_K_M64,000Comfortable (8.4 GB of 12 GB)qwen3.5:9b
Qwen 3.6 35B-A3BQ4_K_M64,000With CPU offload (MoE) (23.3 GB of 12 GB)qwen3.6:35b-a3b
16 GB GPUQwen 3.5 9BQ4_K_M64,000Comfortable (8.4 GB of 16 GB)qwen3.5:9b
Qwen 3.6 35B-A3BQ4_K_M64,000With CPU offload (MoE) (23.3 GB of 16 GB)qwen3.6:35b-a3b
24 GB GPUQwen3.8 27BQ4_K_M64,000Tight (21.9 GB of 24 GB)qwen3.8:27b
Qwen 3.6 27BQ4_K_M64,000Tight (21.9 GB of 24 GB)qwen3.6:27b
Qwen 3.6 35B-A3BQ4_K_M65,536Tight (23.3 GB of 24 GB)qwen3.6:35b-a3b
32 GB GPUQwen3.8 27BQ4_K_M64,000Comfortable (21.9 GB of 32 GB)qwen3.8:27b
Qwen 3.6 27BQ4_K_M64,000Comfortable (21.9 GB of 32 GB)qwen3.6:27b
Qwen 3.6 35B-A3BQ4_K_M65,536Comfortable (23.3 GB of 32 GB)qwen3.6:35b-a3b
Qwen3-Coder 30B-A3B (MoE)Q4_K_M65,536Comfortable (25.7 GB of 32 GB)qwen3-coder:30b
Devstral Small 2 24BQ4_K_M64,000Comfortable (25.8 GB of 32 GB)devstral-small-2:24b
Mac, 64 GBQwen 3.6 35B-A3BQ8_0131,072Comfortable (40.7 GB of 48 GB)—

Your exact machine: filter the matrix or check any model in the analyzer.

Common problems with local models

  • Ollama's Codex guide asks for a context window of at least 64k tokens; LM Studio's guide says more than ~25k. Codex sends a large system prompt, so stay at the higher figure.
  • If oss_provider is unset and you pass no --local-provider, the interactive CLI prompts you to choose, and `codex exec` exits with an error.
  • Custom [model_providers.<id>] entries support only wire_api = "responses", so a server must speak the OpenAI Responses API; the ids ollama and lmstudio are reserved.
  • When launched through `ollama launch codex`, the /model picker lists only installed, tool-capable models; pull extra models before launching.

Other coding agents

How they rank for local models: the best agent harness for local LLMs.

Sources (7), checked 2026-10-20