MultiDex is a Codex-compatible coding agent that runs locally on your computer.

If you want Codex in your code editor (VS Code, Cursor, Windsurf), install in your IDE.
If you want the desktop app experience, run codex app or visit the Codex App page.
If you are looking for the cloud-based agent from OpenAI, Codex Web, go to chatgpt.com/codex.

---

Quickstart

Installing and running MultiDex

Linux x86_64 and ARM64 (including NVIDIA GB10 Spark), plus macOS ARM64, releases are supported. The installer selects the package for the current platform and architecture, installs the codex command and its Code Mode companion, and leaves the existing CODEX_HOME auth, configuration, and session files in place:

curl -fsSL https://github.com/ModelCloud/MultiDex/releases/latest/download/install-multidex.sh | sh

To pin a release, pass --release VERSION to the downloaded installer. Each upgrade keeps prior versioned binaries under CODEX_HOME/packages/standalone/ so rollback can switch the current symlink back to the previous release. Local model endpoints are not contacted during installation or upgrade.

To roll back, inspect the retained release directories and repoint current:

CODEX_HOME_DIR="${CODEX_HOME:-$HOME/.codex}"
ls "$CODEX_HOME_DIR/packages/standalone/releases"
ln -sfn "$CODEX_HOME_DIR/packages/standalone/releases/<previous-release>" \
  "$CODEX_HOME_DIR/packages/standalone/current"

Configure named OpenAI-compatible providers

OpenRouter is built in: set experimental_bearer_token under [model_providers.openrouter] in CODEX_HOME/config.toml and select deepseek/deepseek-v4.1-flash-1017. See OpenRouter configuration for Wafer-only routing. Both the local and OpenRouter DS v4.1 Flash entries default to Max reasoning and use the DeepSeek Harness prompt adapted to MultiDex's tools.

MultiDex keeps the built-in openai provider and its normal ChatGPT/API-key authentication. Add custom providers by name in CODEX_HOME/config.toml; each provider has its own endpoint and bearer-token environment variable:

model = "gpt-6-sol"
model_provider = "openai"

[model_providers.dsv41]
name = "DSV4.1 endpoint"
base_url = "http://10.0.13.33:2120/v1"
env_key = "DSV41_BEARER_TOKEN"
wire_api = "responses"
requires_openai_auth = false

[profiles.dsv41]
model = "DS v4.1 Flash 1017"
model_provider = "dsv41"

Set DSV41_BEARER_TOKEN in the environment before launching MultiDex, then select the dsv41 profile with codex --profile dsv41. The bearer value stays out of TOML, command-line arguments, and MultiDex logs. Repeat the provider and profile sections with distinct names for additional endpoints. Do not set a global openai_base_url when you need official OpenAI models and custom models to share the same session.

For Omnigent, put provider definitions in ~/.local/share/multidex/config.toml and model-to-provider mappings in ~/.local/share/multidex/models.toml. This keeps the registrations out of Omnigent's global provider catalog while letting one MultiDex harness offer official and custom models together:

# ~/.local/share/multidex/config.toml
[model_providers.dsv41]
name = "DSV4.1 endpoint"
base_url = "http://10.0.13.33:2120/v1"
env_key = "DSV41_BEARER_TOKEN"
wire_api = "responses"
requires_openai_auth = false

[model_providers.lab]
name = "Lab endpoint"
base_url = "https://models.example/v1"
env_key = "LAB_BEARER_TOKEN"
wire_api = "responses"
requires_openai_auth = false

# ~/.local/share/multidex/models.toml
[models."DS v4.1 Flash 1017"]
provider = "dsv41"
display_name = "DS v4.1 Flash 1017"
discover_capabilities = true

[models."lab/coder-32b"]
provider = "lab"
display_name = "Lab Coder 32B"

Export the named bearer variables in the Omnigent host service environment. Only models with their configured bearer variable present are offered as routable; selecting one routes to its named provider. Official Codex models continue to use the normal ChatGPT login and OpenAI endpoints. Installation and upgrade never contact or validate a model endpoint.

An orchestrator attaching a TUI with --remote can set MULTIDEX_SKIP_RESUME_GOAL_PROMPT=1 to keep the composer available when resuming a paused, blocked, or usage-limited goal. This only hides the resume-choice modal; it does not resume the goal. Interactive remote connections still prompt by default, and local sessions ignore this flag.

If an endpoint returns response.incomplete with reason max_output_tokens, it has reached its output budget. MultiDex reports that reason immediately and stops the request without retrying or switching compaction models. Increase the inference server's output-token limit or request a shorter response before trying again; increasing connection retries will not raise the output budget.

Switching models and providers

In /model, DeepSeek v4.1 Flash appears once. Select its provider (Local, OpenRouter, or LithosAI), then a variant when available (LithosAI Regular, Fast, or Ultra), and finally the reasoning effort. Models available from only one provider stay in the flat list.

The model picker routes the bundled OpenAI, MultiDex, OpenRouter, and Lithos models to their providers without discarding conversation history. The selected model's context budget applies to the next turn. Moving to a smaller window can require compaction before sampling; moving to a larger window does not by itself require compaction. A model's compaction-compatibility hash can independently require a fresh summary.

After a cold resume, an endpoint transition without verified previous-provider state uses the current model/provider for required compaction instead of sending the previous endpoint's model ID to the new provider.

To check routing, credentials, retained history, and runtime context changes with fake keys and loopback endpoints:

python3 scripts/test_lithos_provider_handoff.py --binary /path/to/multidex

Model at capacity (server_is_overloaded)

When the provider reports the selected model is at capacity ("Selected model is at capacity. Please try a different model."), MultiDex waits and resumes the request automatically — no user interaction and no retry budget is consumed; the request retries indefinitely until capacity frees up. A Retry-After hint from the provider is honored; otherwise the delay defaults to 1 second and is configurable per provider or globally:

[model_providers.dsv41]
# ... other fields ...
server_overloaded_retry_delay_ms = 2000

# Global override (also covers the built-in openai provider)
export CODEX_SERVER_OVERLOADED_RETRY_DELAY_MS=2000

You can also go to the latest GitHub Release and download the appropriate binary for your platform.

Each GitHub Release contains many executables, but in practice, you likely want one of these:

• macOS
  • Apple Silicon/arm64: codex-aarch64-apple-darwin.tar.gz
  • x86_64 (older Mac hardware): codex-x86_64-apple-darwin.tar.gz
• Linux
  • x86_64: codex-x86_64-unknown-linux-musl.tar.gz
  • arm64: codex-aarch64-unknown-linux-musl.tar.gz

Each archive contains a single entry with the platform baked into the name (e.g., codex-x86_64-unknown-linux-musl), so you likely want to rename it to codex after extracting it.

Using Codex with your ChatGPT plan

Run codex and select Sign in with ChatGPT. We recommend signing into your ChatGPT account to use Codex as part of your Plus, Pro, Business, Edu, or Enterprise plan. Learn more about what's included in your ChatGPT plan.

You can also use Codex with an API key, but this requires additional setup.

Docs

• Codex Documentation
• Contributing
• Installing & building
• Open source fund

This repository is licensed under the Apache-2.0 License.