Codex and Creator OS: Post From the OpenAI Coding Agent

You can run your social accounts from the same terminal window where Codex writes code. Codex and Creator OS: post from the OpenAI coding agent is not a metaphor. Codex can call MCP tools, and Creator OS runs a hosted MCP server that exposes tools like create_post, list_posts and get_best_time_to_post. This post shows the actual wiring, one prompt you can copy, and the failure cases you should handle before you let an agent publish to a live account.


  +-------------+        MCP over HTTP        +------------------+
  |    Codex    | --------------------------> |  mcp.creatoros.ca |
  |  (agent)    |   Authorization: Bearer     |     /mcp          |
  +-------------+                             +---------+--------+
        |                                               |
        | reads repo files                              | workspace scoped
        v                                               v
  +-------------+                             +------------------+
  | your project|                             | IG / TT / YT / X  |
  | (prompt, md)|                             | LI / FB / Threads |
  +-------------+                             +------------------+

Codex talks to the Creator OS MCP server; Creator OS holds the platform connections.

What Codex Can and Cannot Post On Its Own

Codex is a coding agent. It reads your repository, edits files, runs commands and calls tools you give it. That is a strong base for content work: your captions can live as files in the repo, your product names stay consistent, and a prompt can assemble a post from real source material.

What Codex does not have is a connection to Instagram, TikTok, YouTube, X, LinkedIn, Facebook or Threads. Platform APIs require per-account OAuth, token refresh, media upload rules, and a long list of per-post fields. An agent cannot invent that. It needs a service that already holds the connections and exposes them as tools.

The split is simple. Codex decides what to say and when. Creator OS holds the accounts and applies each platform’s rules when the post goes out.

Get started for $0

Why Codex Needs Creator OS to Reach Social Channels

Creator OS runs a hosted MCP server at https://mcp.creatoros.ca/mcp. Nothing to install on your machine. You add it as a connector, sign in to Creator OS, pick a workspace, and the tools appear.

Two details make this practical for an agent rather than a human clicking buttons:

Tool names are verbs. create_post, update_post, upload_media_from_url, get_post_analytics, list_comments, reply_to_comment, check_caption_length. A prompt can name one directly.
Destructive tools are marked. Deleting a post, deleting a comment, disconnecting an account or deleting an ad is flagged so the AI app asks you first. Codex has to stop and confirm.

There is also a read-only mode. If you connect with read-only scope, the connection only sees read tools, and the API refuses writes from it. That is the right setting for the first session where you just want Codex to look at your analytics.

If you want the broader picture of how the API layer works, the social media API for AI agents post covers it.

Install Creator OS and Create an API Key

Two paths, depending on whether you want the CLI in the loop.

Path A: web connector only

In Claude or ChatGPT you can add the MCP server as a custom connector with the URL. Codex and other MCP clients need the header form instead, which is the next section.

Path B: CLI plus skills

Install the CLI and initialize skills inside your project:

npx @creatoros/cli@latest init

That installs 14 Claude Code skills including post-everywhere, post-shortform, schedule-posts, respond-to-comments and analytics. Run creatoros sync later to update them; edits you made to a file are not overwritten, the update lands as a .new file.

Now check auth and see what is connected:

creatoros auth:check
creatoros accounts:list
creatoros accounts:health

accounts:health is the one people skip. It tells you if a connected account has a broken token before you queue a week of posts against it.

Grab an API key (cos_live_...) from Creator OS. A key is pinned to one workspace, meaning one set of socials. That scoping matters: an agent holding one workspace’s key cannot post to another workspace’s accounts.

Connect Codex to Creator OS in Your Project

Add the server with the HTTP transport and the bearer header. The Claude Code form looks like this:

claude mcp add --transport http creatoros https://mcp.creatoros.ca/mcp \
  --header "Authorization: Bearer cos_live_..."

For Codex and other MCP clients, use the equivalent mcpServers entry with type http, the URL and the same Authorization header. The exact file location differs per client; the docs at creatoros.ca/docs/mcp list the supported apps, including Cursor, Windsurf and VS Code.

Restart the session, then ask Codex to list its tools. If it can see create_post and get_best_time_to_post, the connection is live. If it sees only read tools, you chose read-only scope at sign-in. Reconnect and pick read and write.

Write and Approve a Post With a Prompt

Here is a prompt that uses the tool set directly. Put it in your repo as a file so it is versioned with the code it describes:

Read CHANGELOG.md and pick the three most user visible changes
from this week. Draft one Instagram caption under 180 characters
and one X post under 260 characters. Run check_caption_length on
each draft and fix anything that fails. Show me both drafts and
wait. Do not call create_post until I reply APPROVE.

Three things are happening. Codex grounds the copy in a real file instead of guessing. It calls check_caption_length before you ever see the text, so a 300 character caption does not become a failed publish. And the prompt contains an explicit stop.

The stop matters more than it looks. Approval is a workflow problem, not a syntax problem. If you want the agent to ping you when drafts are ready instead of leaving you watching a terminal, see the notification app post.

When you approve, the follow-up prompt can be short:

Approved. Upload media/release-card.png with upload_media_from_url,
then create_post with both captions. Save the returned post ids to
posts.json. Do not schedule anything yet.

Saving the post ids is the part that pays off later. A post_ id is what you need for analytics, edits and unpublishing.

Publish to Instagram, TikTok, and More

create_post takes a caption and a list of platforms. Creator OS then applies each network’s rules. That is the entire reason to route through it rather than writing six API clients.

Some examples of what gets handled for you:

Instagram. Reels, feed images, carousels and Stories. Custom Reel cover or a frame offset. Up to three collaborators. A firstComment field, because captions have no clickable links. The API cannot attach trending audio, so sound has to be baked into the file.
TikTok. Required consent fields like content_preview_confirmed and express_consent_given, which the simple post form fills in. draft true sends the video to your TikTok Creator Inbox so you can finish it in the app.
YouTube. Long-form and Shorts, with the first comment posted and pinned. Custom thumbnails on long-form, not on Shorts.
Threads. 500 characters per post, multi-post threads via threadItems, and a topic_tag that cannot contain periods or ampersands.

That last one is a good example of why an agent needs a proxy. A topic tag rule with a punctuation restriction is the kind of detail a model will get wrong once and then keep getting wrong. The post options reference lists every field per platform.

For a deeper Instagram walkthrough, including the account type requirement, read how to let an AI agent post to Instagram.

Handle Scheduling, Errors, and Rollbacks

Scheduling is a tool call and a timezone. Ask Codex for a slot from your own data rather than a generic best practices list:

Call get_best_time_to_post for instagram and tiktok, then schedule
the approved drafts into the two highest scoring windows this week.

The reasoning behind using your own numbers instead of a generic chart is in best time to post on Instagram, your own data.

Now the failure modes. When you name platforms in a post and some are not connected, Creator OS returns them in missing_platforms and still publishes to the rest. That is deliberate: a dead LinkedIn token should not block your TikTok. But it means Codex should read the response, not assume success. Tell it explicitly:

After create_post, if missing_platforms is not empty, print the
list and stop. Do not retry automatically.

If the CLI is in the loop, the recovery commands are direct. creatoros posts:retry <postId> for a failed publish, creatoros posts:get <postId> to inspect status, and creatoros posts:unpublish <postId> --platform facebook|twitter|threads|linkedin|youtube when something needs to come down. Text edits after publishing are available on X, Facebook, LinkedIn and YouTube through posts:edit.

Test the whole loop in draft first. Create the post with --draft, confirm Codex handled the media and the caption length, then schedule it for real.

Safety and Rate Limits Before You Go Live

Three controls to set up before Codex touches a live account:

Start read-only. Connect with read-only scope and let Codex run list_posts, get_post_analytics and get_follower_stats for a session. You learn its habits with zero downside.
One workspace per client. API keys are pinned to one workspace, so a client project cannot reach a different brand’s accounts. Use separate keys rather than sharing one.
Keep the confirmation prompts on. Destructive tools are marked so the app asks first. Do not prompt Codex to route around that.

You can also validate before you publish. creatoros validate:post-length --text "..." and creatoros validate:media --url ... catch the two most common hard failures without touching a platform.

If you would rather run a local agent harness than an MCP connector, the open-source Social Agents repo (github.com/kevinbadi/social-agents) runs your socials on Creator OS and starts with npm start creatoros social-agents. It brings its own local dashboard and works with a logged-in Claude Code session, an API key, or any Anthropic-compatible endpoint. The launch walkthrough is on Kevin’s launch video, and there are more builds on the KevBuildsApps YouTube channel.

For agents that need to act inside a browser session rather than through an API, read the Chrome relay post.

Get Started With Codex and Creator OS

The shortest path: create an account, connect one social account, add the MCP server to Codex with your cos_live_ key, and ask the agent to list your posts. Nothing publishes until you tell it to.

The Creator plan is $14.99/month or $59.99/year and covers 11 platforms plus Skool, one account on each. API keys, the MCP server, the CLI and the agent skills are included in every plan, so you are not paying extra for the agent layer. Start at creatoros.ca/sign-up.

When you want the full tool list, per-platform fields and client setup for Cursor, Windsurf and VS Code, go to creatoros.ca/docs.

Get started for $0
Free trial for new subscribers. Cancel anytime.