Connect Codex to Token Factory

Corvex Token Factory exposes an OpenAI Responses API endpoint, which is the protocol Codex requires from a custom model provider. Add Token Factory as a provider in ~/.codex/config.toml, describe its models in a model catalog file so Codex knows their real context windows, and select a model. The Codex CLI and the Codex IDE extension read the same configuration. Verified with Codex 0.162.1, October 2026.

Prerequisites

A Token Factory API key. See Authentication.
Codex CLI 0.162.1 or later (npm install -g @openai/codex) and jq.
A model ID from the Token Factory catalog.

Context window and compaction

Codex treats a model ID it doesn’t know as a 272,000-token model, and model_context_window cannot raise that ceiling. A model catalog file (model_catalog_json) gives Codex each model’s real window. Codex does not send max_output_tokens, so it reserves no output room of its own (max_tokens = 0 in the formula below). Its default compaction point is 90% of the window. A single turn with large tool results can push the context past the compaction point before compaction runs, so leave about 15% headroom: set model_auto_compact_token_limit to (context_window − max_tokens) / 1.15, rounded down:

Model	Context window	Output cap	`max_tokens` sent	Compaction window
`zai-org/GLM-5.3`	393,216	393,216	0	341,926
`deepseek-ai/DeepSeek-V4-Flash-0731`	1,048,576	1,048,576	0	911,805

Codex counts 95% of the window as usable (373,555 tokens for a 393,216-token window); after the first turn, /status shows that usable window. GET /v1/models reports each model’s current context_window and max_output_tokens.

Store the key in your shell

Codex reads the API key from the environment variable named by env_key. Export it in your shell profile so both the CLI and the IDE extension can read it.

export CORVEX_API_KEY="sk-corvex-..."

Create the model catalog

Build the catalog from Codex’s bundled one, so the Token Factory models keep Codex’s standard instructions. The command writes ~/.codex/token-factory-models.json:

create-catalog.sh

codex debug models --bundled | jq '
  (.models | min_by(.priority) | .model_messages) as $messages
  | {models: [
      {slug: "zai-org/GLM-5.3", context_window: 393216},
      {slug: "deepseek-ai/DeepSeek-V4-Flash-0731", context_window: 1048576}
    ] | map(. + {
      display_name: .slug, description: null, max_context_window: .context_window,
      effective_context_window_percent: 95, model_messages: $messages,
      supported_reasoning_levels: [], shell_type: "unified_exec", visibility: "list",
      supported_in_api: true, priority: 99, availability_nux: null, upgrade: null,
      support_verbosity: false, default_verbosity: null, apply_patch_tool_type: null,
      web_search_tool_type: "text", truncation_policy: {mode: "bytes", limit: 10000},
      experimental_supported_tools: []
    })}' > ~/.codex/token-factory-models.json

The catalog replaces Codex’s built-in model list, so it lists only Token Factory models.

Add the provider to config.toml

Create or edit ~/.codex/config.toml:

config.toml

model = "zai-org/GLM-5.3"
model_provider = "corvex"
model_catalog_json = "~/.codex/token-factory-models.json"
model_auto_compact_token_limit = 341926

[model_providers.corvex]
name = "Corvex Token Factory"
base_url = "https://api.tokenfactory.corvex.cloud/v1"
env_key = "CORVEX_API_KEY"
wire_api = "responses"

Enter the model ID exactly as the catalog lists it, including the namespace/ prefix. Do not add a corvex/ prefix.

Add a profile for each other model

The compaction limit is per model, so keep one profile file per model next to config.toml. Codex layers ~/.codex/<name>.config.toml on top of config.toml when you pass --profile <name>:

deepseek.config.toml

model = "deepseek-ai/DeepSeek-V4-Flash-0731"
model_auto_compact_token_limit = 911805

Run Codex

cd /path/to/your/repo
codex                      # zai-org/GLM-5.3
codex --profile deepseek   # deepseek-ai/DeepSeek-V4-Flash-0731

To run a single task non-interactively:

codex exec "Summarize what this repository does."

Verify the connection

Ask Codex to read a file in the repository, then run /status. It shows Corvex Token Factory, the model ID, and a usable window of 95% of the model’s context window. A window of 258,400 means Codex did not load the catalog; check the model_catalog_json path.

How Token Factory serves the Responses API

POST /v1/responses accepts the Codex request shape, including instructions, input items, tools, tool_choice, max_output_tokens, temperature, and stream. The gateway translates each request to the model’s Chat Completions interface and returns Responses-shaped output, with streaming events in the Responses event order. The model’s reasoning is returned as a reasoning output item ahead of the message item. Conversation persistence is not exposed: store and previous_response_id have no effect, and there is no /v1/conversations endpoint. Codex resends the full conversation on each turn, so this does not affect normal use. Request fields that control OpenAI-hosted reasoning (reasoning, include) are accepted and ignored; the model reasons with its default settings. See the API reference for the documented request and response fields.

Compatibility notes

Context length. If a session exceeds the model’s window, the gateway returns context_length_exceeded. Run /compact in Codex, or start a new session. See Errors.
Output budget. Reasoning tokens count toward the output budget. If a response is truncated, shorten the prompt or split the task.
Input types. Neither model accepts image input. Do not attach images with -i; the request returns a structured error.

Use the OpenAI SDK — Chat Completions from Python or TypeScript.
Connect OpenCode — another terminal coding agent on the OpenAI-compatible surface.
Errors — authentication, rate-limit, and validation responses.

Get Started

Models & Pricing

Integrations

API Reference

Security & Compliance

Reference

Prerequisites

Context window and compaction

How Token Factory serves the Responses API

Compatibility notes

​Prerequisites

​Context window and compaction

​How Token Factory serves the Responses API

​Compatibility notes

​Related documentation

Prerequisites

Context window and compaction

How Token Factory serves the Responses API

Compatibility notes

Related documentation