Hermoso is a marketing MCP server: 922+ tools over one OAuth’d endpoint that let an agent research the ads already running in a market, generate finished image and video ads, publish them, and build the paid campaigns behind them. Codex reaches all of it through one connection. This page covers one slice of that: what Codex can actually do on LinkedIn, and what LinkedIn will not let it do.
Connect Hermoso to Codex
- One command, and no key.
Codex reads the marketplace file in Hermoso’s public repository and installs its five skills (research, ad-from-brand, generate, product-photoshoot and marketing). The skills run thecodex plugin marketplace add hermoso-ai/hermoso && codex plugin add hermoso@hermosohermosocommand throughnpx, so no tool list is loaded into your sessions and nothing costs context until a command runs. - Sign in once.
It opens a browser sign-in and stores an owner-only credential innpx -y hermoso auth login~/.hermoso/config.json. No browser on that machine, say CI or a container? Create a key under MCP & CLI → Terminal & API keys and runnpx -y hermoso auth login --token hmk_…. A key carries real spend authority, so treat it like a password. - Optional: the MCP server as well.
That writes the server intocodex mcp add hermoso -- npx -y hermoso mcp~/.codex/config.tomland reads the same stored credential, so no key goes in the file. Prefer the hosted endpoint?codex mcp add hermoso --url https://app.hermoso.ai/mcp, thencodex mcp login hermosofor the OAuth sign-in. The skills reach every tool without either. - Ask Codex for something. “Pull my three closest competitors’ longest-running ads and remix the best hook into a 9:16 video,” or “post this render to the channel.” The agent picks the tools; you approve the spend.
Codex keeps MCP servers in ~/.codex/config.toml, or .codex/config.toml in a trusted project, and that file is TOML, not JSON: a copied mcpServers block will not load, and a hand-written entry is [mcp_servers.hermoso] with command = "npx" and args = ["-y", "hermoso", "mcp"]. The trailing mcp is load-bearing: npx -y hermoso with no subcommand prints help and exits. Codex reads plugin marketplaces in the same .claude-plugin layout Claude Code uses, which is why one repository installs in both.
Then connect LinkedIn itself, once, under Settings → Connectors in Hermoso. That is an OAuth consent screen, so it is the one step that needs a browser. Everything after it is a tool call.
The LinkedIn tools Codex gets
post_to_linkedin_page: publishes text, an image, a video or a multi-image post to a company Page, with alt text.post_to_linkedin: publishes to a personal profile, which stays switched off until you explicitly share that identity.list_linkedin_pages: the Pages you have shared with this brand, with their organization ids.schedule_post: queues it, carryinglinkedinOrganizationIdso the right identity posts.manage_linkedin_post: edit the copy of a post, or delete it.linkedin_page_analytics: organic Page performance: follower gains, Page views, and post impressions, clicks and engagement, per Page or per post.search_linkedin_ads: the LinkedIn Ad Library, the one public ad library that exposes real impression counts.
LinkedIn is the channel where identity is a permission rather than a preference. One person usually administers several companies’ Pages, so Hermoso makes the postable identities an explicit choice: only the Pages you have ticked under Manage accounts can be posted to, an unticked Page is refused server-side rather than in the UI alone, and the personal profile is off unless you turn it on. An agent that asks which Page rather than assuming is doing the right thing.
LinkedIn limits worth knowing before you automate anything
| Text | A LinkedIn post is capped at 3,000 characters. It is enforced when the post is created or queued, so an overlong post is refused while you are still there to fix it instead of being silently cut on the way out. |
|---|---|
| Visibility | Public only. A LinkedIn post publishes to the profile or the company Page, and “LinkedIn has no unlisted, private or draft post.” |
| Carousel | 2 to 20 images, from a company Page. Both bounds are LinkedIn’s own: “A MultiImage post is a post containing multiple images (minimum of 2 images and maximum of 20 images)”, and the schema adds “Only supports images”. Two refusals follow from that and both are by name rather than by silent downgrade: a video slide is refused because LinkedIn’s multi-image post cannot carry one, and a personal-profile carousel is refused with the way forward: pass linkedinOrganizationId (from list_linkedin_pages), because “a personal-profile carousel is not something Hermoso can publish.” |
| Which account | linkedinOrganizationId. Unset means the member’s own profile, never “probably the company”. The field is how you say which Page, and leaving it out is a decision rather than an accident. |
Try it in Codex
Ask for it in ordinary language. These are plain sentences, not a DSL:
“Non-interactive: publish the weekly digest to the company Page with the four charts as a carousel and alt text on each, then print the post URL.”
That walks list_linkedin_pages → plan_ad → render_ad → post_to_linkedin_page → linkedin_page_analytics. Interactively Codex shows the call before it runs it. Non-interactively there is nobody to ask, which is worth designing around on any step that publishes or spends.
Codex on LinkedIn: the part that bites
Resolve list_linkedin_pages once and store the organization id in the job config. Leaving it unset does not mean “work it out”. It means the personal profile, which is off until explicitly shared, so an unattended run would fail on a permission rather than post to the wrong place. That fail-closed default is the right one; just do not rely on discovering it at 9am.
For unattended runs, prefer the hermoso CLI with --json over the full tool manifest. It is far cheaper in context and easier to assert on.
Research LinkedIn before you post
search_linkedin_ads reads the LinkedIn Ad Library, which is unusual among public ad libraries in exposing real impression counts, so B2B competitor research on LinkedIn can be ranked by actual delivery rather than by guesswork about what is running.
The paid half
create_linkedin_ads_campaign_group → create_linkedin_ads_campaign → create_linkedin_ads_creative builds the tree, and everything is born as a draft. set_linkedin_ads_status is the only switch that arms real money and it refuses without an explicit confirmation. One access note that is better said than discovered: LinkedIn meters ad writes by the API tier your own ad account holds. Reporting and reads are unrestricted, and on the development tier an app may edit a small number of ad accounts, each registered by hand.
What Codex cannot do on its own
An agent with a budget is only useful if the blast radius is bounded, so the fence is in the server rather than in a prompt. Every ad campaign, ad set, ad group and ad is created paused, and the status switches that arm real money refuse to run without an explicit confirmation flag. Every render is quoted first (hermoso_capabilities publishes each model’s exact credit cost before anything spends). Credits are reserved before the first provider call and settled at the exact cost afterwards, so a failed dispatch refunds rather than leaving you billed for nothing. Deletes are confirm-gated too, and every campaign tree is read back from the ad platform before your agent tells you what it built.
One step is deliberately not headless: linking an account is an OAuth consent screen, so you connect the channel once in a browser. Everything after that is a tool call.
The rest of the matrix
From Codex, post to: Facebook · Instagram · Threads · TikTok · YouTube · X · Pinterest · Bluesky · Telegram. Or see everything Codex can publish to.
Post to LinkedIn from: Claude · ChatGPT · Claude Code · Cursor · Cline · OpenClaw · Grok Bot · Hermes · Gemini CLI · Perplexity · Le Chat · Claude Cowork · Perplexity Computer · Paperclip · Ads from the command line
Frequently asked questions
Can Codex post to LinkedIn?
Yes, with Hermoso's Codex plugin. Run codex plugin marketplace add hermoso-ai/hermoso && codex plugin add hermoso@hermoso, which installs Hermoso's skills, then sign in once with npx -y hermoso auth login. The sign-in opens a browser and stores the credential itself, so there is no key to create and none to paste. For the MCP server as well, run codex mcp add hermoso -- npx -y hermoso mcp. Codex then has 922+ tools, including the ones that publish to LinkedIn, plus the ones that research the ads already running in your market and render the creative in the first place.
How do I connect LinkedIn to Codex?
Two steps, once each. Run codex plugin marketplace add hermoso-ai/hermoso && codex plugin add hermoso@hermoso, which installs Hermoso's skills, then sign in once with npx -y hermoso auth login. The sign-in opens a browser and stores the credential itself, so there is no key to create and none to paste. For the MCP server as well, run codex mcp add hermoso -- npx -y hermoso mcp. Then connect LinkedIn inside Hermoso under Settings → Connectors, which is an OAuth consent screen in a browser. After that Codex publishes to LinkedIn as an ordinary tool call.
Can Codex post privately or as a draft to LinkedIn?
No. LinkedIn has no unlisted, private or draft post. A post publishes to the profile or the company Page. Hermoso refuses a visibility LinkedIn cannot honour at the moment you ask for it, rather than quietly publishing something weaker.
Can Codex schedule a LinkedIn post for later?
Yes. schedule_post queues it on one calendar shared with every other channel, with a per-channel caption if the copy should differ, and list_scheduled and cancel_scheduled read and pull it back. LinkedIn is one of the eleven channels that queue.
What is the LinkedIn caption limit when posting from Codex?
A LinkedIn post is capped at 3,000 characters. Hermoso enforces it when the post is created or queued, so Codex is told immediately rather than failing later when nobody is watching.
What does it cost to post to LinkedIn from Codex?
Publishing to LinkedIn costs no credits; LinkedIn charges nothing per API call. Credits are spent on the render. Plans are $29, $69 and $149 a month for 1,200, 3,000 and 7,000 credits, and every feature is on every plan including Free. The only exception is enterprise SSO.
Start free with 30 credits at signup and 250+ more to earn. Every feature is on every plan. Credits are spent only on AI models and Ad Spy research; posting, scheduling, ads management and analytics are free.
Start free → See pricing