How it works

Start with an assistant you already use. Give it account context, not your Threads credential, and decide what it may publish.

1. Choose your assistant

   Select an exact supported client and environment. Codex and Claude Code have recorded read-only checks. ChatGPT desktop has an official configuration path that still needs an end-to-end gateway validation; ChatGPT web requires a plugin we have not published.

2. Connect your Threads account

   Meta operates the account sign-in. The resulting Threads credential is encrypted by the gateway and is never shown to the assistant.

3. Start your assistant read-only

   Its separate connection can inspect only the read capabilities you grant on that account. It cannot publish a post or reply until a workspace owner deliberately adds the matching capability.

4. Ask for a useful account review

   The first success is not "a key was created." The assistant should inspect its effective permissions, then read recent posts, replies, or available results and return something useful.

5. Let it prepare the next post or reply

   The assistant drafts in the conversation where you already work. If it has proposal capability, the publishable action enters the gateway with a reason and an idempotency key.

6. Choose review-first or automatic publishing

   Approval-required is the safe default. When you deliberately select Automatic publishing, the assistant can publish without a per-item approval while its granted capabilities, posting limits, prohibited content rules, duplicate-reply protection, and Meta quota still apply.

7. Bring the observations into the next session

   Available results are stored at fixed post-age horizons. Comparisons carry medians, sample sizes, missing-data behavior, and the statement that an observed association does not establish cause.