Does Codex support remote MCP servers like this?

Yes — Codex loads MCP servers from ~/.codex/config.toml, and an HTTP server entry with a url plus http_headers is all NinjaChat needs. No local process, no npx wrapper.

How do I keep an autonomous agent from overspending?

The nj_sk_ key carries a monthly spending limit enforced server-side — a runaway loop stops at the cap, not at your card. Failed generations auto-refund and requests are idempotent, so retries are free of double-billing.

Which image model should Codex default to?

It doesn't have to guess — search_models picks by task and price. As a rule of thumb: z-image-turbo ($0.01) for drafts and bulk, nano-banana-2 ($0.067) for balanced quality, gpt-image-2 ($0.128) when text rendering inside the image matters.

Can Codex edit images from my repo, not just generate new ones?

Yes — upload_asset pushes a local file to NinjaChat, then edit_image applies a prompt-based edit (qwen-image-edit-plus, $0.03). The result comes back as a URL Codex downloads into the workspace.

Is the same key usable across Cursor, Claude Code, and VS Code?

Yes — one nj_sk_ key authenticates any header-auth MCP client. If you want per-tool accounting or tighter CI limits, create separate keys; each has its own spending limit.

What happens if two parallel Codex tasks fire the same generation?

Idempotency keys ensure a repeated request doesn't bill twice, and each distinct generation is priced per output. get_account_balance gives the live headroom if you want the agent to gate itself.

Create free key

Codex

Codex, meet every creative model.

That ships the mark.

Three lines of config.toml give OpenAI's Codex agent real image and video generation — 22 models, per-output pricing, and cost estimates it can check before it spends.

Connect Codex

Step 1

Connect Codex. Once.

Get your API key.

Key creation is free. You’ll need developer credit to activate it.

What happens after I get my key?

We’ll show you how to add NinjaChat to Codex, with your key already filled in. Then you can send your first prompt.

Check your connection Free

Enable NinjaChat in a new chat, then send this check.

View check prompt

Use NinjaChat to check my account balance and spending headroom, then list the available image models. Do not generate anything or add credits. If NinjaChat is unavailable, tell me which setup step is missing.

Balance & spending limit

Different client?Cursor Claude ChatGPT Claude Code VS Code Hermes OpenClaw

Step 2 · First ask in Codex

Paste this.
Review the quote.

Check the connection above, then paste your request. Approve the quoted cost before generating.

Example style · your result will vary

Say this in Codex

Use NinjaChat for this task: Make one dark cinematic still of a glass bottle on wet stone. After approval, save it under public/ninjachat-hero with the correct file extension and verify the file. First check my balance and use estimate_cost to quote the model, quantity, dimensions or duration, and total price. Ask for any missing reference files. Do not generate, publish, or add credits until I approve the quote. For video, retain the job ID and poll get_generation_status until complete; resume that job instead of starting another. Return the finished media and download links.

Back to connect

Step 3 · What Codex paints

KIYO on the mark. Header updated.

Four bottle directions on recraft. Keep the strongest. Patch the header.

KIYO persimmon wall — **KIYO**02 Persimmon

Four directions. One bottle that holds the brand.

Models

Every model Codex can paint with.

Image and video engines behind one URL — Codex asks, NinjaChat runs them. New models appear here; your connection never changes.

Google Veo 3.1Veo 3.1

Seedance 2.0Seedance

An editorial shot of a cobalt coat in white dunes — a real FLUX.2 Pro run

FLUX.2 ProFLUX.2 Pro

Kling 2.6 ProKling

Google Veo 3.1 FastVeo Fast

A flamenco dancer mid-turn under a single spotlight — a real Seedream run

Seedream 4.5Seedream

A red-sailed boat crossing a mirror-still turquoise fjord — a real Nano Banana 2 run

Google Nano Banana 2Nano Banana 2

Seedance LiteSeedance Lite

A matte espresso machine in a shaft of morning light — a real Nano Banana Pro run

Google Nano Banana ProNB Pro

A red fox mid-leap over a frozen creek at sunrise — a real Nano Banana run

Google Nano BananaNano Banana

A hummingbird frozen mid-air at a fuchsia flower — a real Z-Image Turbo run

Z-Image TurboZ-Image

GPT Image 2's take on the tiny astronaut asleep on a crescent moon

GPT Image 2GPT Image

What Codex can call.

Make

generate_imageCreate images and download the full-resolution results.
edit_imageChange a background, color, or detail with a prompt.
get_generation_statusSee whether a video is ready to view.

Decide

search_modelsCheck model capabilities and prices.

Money

estimate_costCheck the price before generating.

Also on the server

create_mediagenerate_speechgenerate_videoimage_to_videoedit_videoextend_videosearch_creationsget_creationsave_creation_aslist_creative_entitiescompare_modelsupload_assetprepare_scrub_videopublish_pageget_account_balanceadd_credits

Troubleshooting

When it doesn't connect.

Codex starts without the ninjachat tools

TOML is picky: the section must be [mcp_servers.ninjachat] (underscore, not dash) in ~/.codex/config.toml, and http_headers is an inline table. Restart the session after editing — config loads at launch.

401 errors on every tool call

The Bearer key is invalid or was truncated in the TOML string. Regenerate at ninjachat.ai → Developers → Keys and paste the full nj_sk_ value inside the quotes.

Requests hang in restricted-network mode

Codex sandboxes network access. Allow requests to www.ninjachat.ai in your Codex network policy, or run the task with network enabled.

Video results never arrive

Video is async — the first call returns a job, and get_generation_status fetches the finished file a minute or two later. Prompt Codex to poll before concluding the job failed.