Blog ·

How to add agent skills to ChatGPT

ChatGPT doesn’t have a built-in “skills” feature the way Claude does — but it does support full MCP custom connectors in Developer Mode. Point ChatGPT at a Git-backed skill library’s MCP endpoint and it can search and load your governed agent skills on demand. You don’t author skills in ChatGPT; you serve them to it.

First, the plan reality check

Custom MCP connectors require Developer Mode, which is available on the paid tiers — ChatGPT Plus, Pro, Business, or Enterprise. The free tier does not support custom connectors. The good news: you do not need Enterprise. Any paid plan can connect to a skill library, so this works for individuals and small teams, not just large orgs.

One caveat ChatGPT’s own docs flag: custom connectors aren’t verified by OpenAI, so only connect a server you trust and control.

Why a skill library, not copy-paste prompts

You could paste instructions into a ChatGPT custom GPT, but that doesn’t scale: every update is manual, nothing is versioned, and the same knowledge has to be re-pasted into Claude, Codex, and every other assistant. A skill library fixes that by being the single source the connector reads from:

  • Git is the source of truth — one canonical, versioned copy with history and review.
  • Governed access — org / team / user permissions decide which skills each person’s ChatGPT can see.
  • Portable — the same SKILL.md works in Claude, Codex, and ChatGPT, so you maintain it once.

Step by step: connect ChatGPT

  1. In ChatGPT, open your profile → Settings Connectors (or Apps & Connectors).
  2. Click Advanced at the bottom and toggle Developer mode on.
  3. Choose Add new connector and give it a name and description (e.g. “Skills Manager”).
  4. Enter the library’s MCP server URL. For Skills Manager that’s https://www.skillsmanager.uk/mcp.
  5. Set authentication to OAuth, sign in, and confirm I trust this application.
  6. Save. ChatGPT can now discover the skills you can access.

Make ChatGPT actually load your skills

A connected library is necessary but not sufficient — the model still decides when to reach for a skill. Two things make loading reliable:

  • Favourites. Star the skills you always want available (admins can enforce team favourites) so there’s a short, preloaded list.
  • An enforce-loading instruction. Add a one-line directive to ChatGPT’s custom instructions telling it to pull the favourites list at the start of every session. The exact copy-paste text is on the Enforce Skill Loading page.

Recap

RequirementDetail
PlanAny paid tier (Plus / Pro / Business / Enterprise)
ModeDeveloper Mode → custom MCP connector
ServerRemote MCP URL (not a local stdio server)
AuthOAuth to your skill library
Reliable loadingFavourites + enforce-loading instruction

Related: create skills in Claude that work in ChatGPT, sync skills between ChatGPT and Claude, and connecting your assistants.

Bring your governed skills into ChatGPT. Start free and connect in minutes.

Last updated: June 15, 2026

Your skills, now in ChatGPT