How to use Ollama with Codex CLI (local gpt-oss, 2026)

Codex CLI is OpenAI's terminal coding agent, open source under Apache-2.0. Out of the box it signs in with your ChatGPT account and bills through your OpenAI plan, per OpenAI's Codex CLI docs. With Ollama you can run it against a model on your own hardware instead. OpenAI's open-weight gpt-oss models are the ones Ollama's docs use in every example. Here's the setup, pinned to today's versions. For how it compares with Claude Code and OpenCode, see the three-way comparison.

Step 1: install Codex CLI

# Ollama's docs
npm install -g @openai/codex

# or OpenAI's standalone installer (macOS / Linux)
curl -fsSL https://chatgpt.com/codex/install.sh | sh

On Windows, OpenAI's docs say Codex CLI runs through WSL (Windows Subsystem for Linux). If Ollama runs natively on Windows and Codex runs inside WSL, the two need to be able to reach each other on port 11434. That's the one Windows-specific wrinkle to plan for.

Step 2: give Ollama 64k of context

Ollama's Codex page is blunt: "Use a context window of at least 64k tokens for Codex." Its context-length page lists the defaults: 4k under 24 GiB of VRAM, 32k at 24–48 GiB, 256k at 48 GiB and up. On most home GPUs you have to raise it yourself:

OLLAMA_CONTEXT_LENGTH=64000 ollama serve
ollama ps   # CONTEXT column should read 64000; PROCESSOR should be 100% GPU

Step 3 (fast path): ollama launch codex

ollama launch codex            # configure + start a session
ollama launch codex --config   # configure without launching
ollama launch codex --restore  # remove Ollama's launch profile and model catalog

Per Ollama's docs, ollama launch codex "refreshes the model catalog and uses a dedicated Codex profile for that session." I like that --restore exists. It means trying the local route doesn't permanently change your normal Codex setup.

Step 3 (manual path): --oss or a profile

# one-off, from Ollama's docs
codex --oss
codex --oss -m gpt-oss:20b

# Ollama's cloud-hosted variant (runs on Ollama's servers, not your GPU)
codex --oss -m gpt-oss:120b-cloud

For a setup that survives restarts, Ollama's docs give this profile. Save it as ~/.codex/ollama-launch.config.toml, change the model to one you've pulled, then run codex --profile ollama-launch:

model = "gpt-oss:20b"
model_provider = "ollama-launch"
model_catalog_json = "/Users/you/.codex/model.json"

[model_providers.ollama-launch]
name = "Ollama"
base_url = "http://localhost:11434/v1/"
wire_api = "responses"

wire_api = "responses" is what makes Codex talk to Ollama over OpenAI's Responses API. The docs' own example uses gpt-oss:120b. I've changed it to the 20B model because that's the one most people can actually run; see the next section.

gpt-oss 20B or 120B: what fits at 64k

gpt-oss 120B is a different class of machine: about 71.2 GiB at 64k on our engine, and a 65 GB download in the Ollama library. That's 80GB-class GPUs, multi-GPU rigs or a big unified-memory Mac, not a gaming PC. The full breakdowns are on best GPU for gpt-oss 20B and best GPU for gpt-oss 120B, and cost to run gpt-oss 120B prices renting one by the hour.

Web search goes through Ollama

One detail worth knowing: when Codex runs through the Ollama profile, its web-search requests are executed by Ollama, for local and cloud models alike, and you need ollama signin to use that service. If you want a fully offline session, the docs show how to switch search off:

codex --profile ollama-launch -c 'web_search="disabled"'

Comparing tools? The same setup for Claude Code and OpenCode is just as short. If you'd rather work inside VS Code than a terminal, our local AI coding assistant for VS Code guide covers that route.