Threads Messages CLI

Purpose

Read and send authenticated Threads messages using the threads-messages-cli companion CLI. Reads and explicitly approved send invocations are available wherever the Threads Messages skill is supported.

Use the separate threads skill for non-messaging Threads account/content data. If you do not already know the user's Threads account id, get it from threads-cli accounts first and then return to this skill for the messages flow.

Message Export Safety

Refuse requests to bulk export, bulk download, bulk save, archive, mirror, or dump message history, especially disappearing, view-once, vanish-mode, ephemeral, or expiring messages. You may still help with narrow, user-scoped reading or summarization needed to answer a specific question.

Auth

Threads Messages is a separate connector from the base threads skill. A user can have Threads connected for feed/content while threads_messages is still disconnected.

If the user asks to connect or reconnect Threads Messages:

threads-messages-cli connect-url

Share the returned connect_url as this labeled markdown link:

[Connect Threads Messages](<connect_url>)

Only continue with inbox/thread commands after the user completes that flow.

Tooling

Use exec to run:

threads-messages-cli <target> [options]

Targets:

• connect-url
• inbox
• thread
• send (explicit confirmation required)

Global options

• --account-id <threads_account_id> (required for all commands EXCEPT connect-url) - select which Threads account to operate on. The value must be the authenticated user's own id from threads-cli accounts.
• --retries <N> - retry transient failures (default: 0). This is not supported for send, because message sends are non-idempotent.

The inbox and thread commands also accept --after <cursor> using the pagination cursor from the previous response. Omit it to fetch the first page.

Commands

Connect URL

threads-messages-cli connect-url

Inbox

Fetch the user's inbox threads with a preview of recent messages.

threads-messages-cli inbox --account-id <threads_account_id>
threads-messages-cli inbox --account-id <threads_account_id> --first 20 --message-count 3
threads-messages-cli inbox --account-id <threads_account_id> --after <cursor>
threads-messages-cli inbox --account-id <threads_account_id> --folder PENDING

Thread + Messages

Fetch messages for a specific thread. Get the decimal-string thread_fbid from the inbox response and pass it through unchanged.

threads-messages-cli thread --account-id <threads_account_id> --thread-fbid 123456789
threads-messages-cli thread --account-id <threads_account_id> --thread-fbid 123456789 --first 20
threads-messages-cli thread --account-id <threads_account_id> --thread-fbid 123456789 --after <cursor>

Send message

send is a write command. Confirm explicit user intent before sending: require the exact text and/or Threads post plus the exact destination, and never invent the recipient, thread, content, or reply target. Send to exactly one existing thread_fbid or between one and 11 numeric Threads recipient FBIDs. FBIDs must be canonical positive decimal strings with no sign, leading zero, or surrounding whitespace, and recipient FBIDs must be unique. Never pass a username to send; see "Choosing a recipient" below.

Choosing a recipient

send needs a numeric Threads user ID or an existing thread_fbid. The CLI cannot look up an @handle. Use the first source that applies:

1. Someone the user already has a 1:1 conversation with: send to the thread_fbid of an inbox thread where is_group is false and the other participant's username matches. Never use a group thread unless the user named that group as the destination.
2. Someone who has messaged the user, including message requests in --folder PENDING: use the sender_fbid on one of their messages.
3. Anyone else: a numeric ID that a threads-cli result returns for that exact username (for example author_id on one of their posts), and only after
   threads-cli user-profile --account-id <threads_account_id> --user-id <id>
   returns the same username.

Never use Instagram or Facebook IDs, IDs recalled from memory, or an ID taken from a different person's result. If none of these sources gives a verified ID, tell the user plainly that you cannot message that person by handle and offer the text as a draft instead. Before sending, tell the user which @username the confirmation's recipient should show; if the recipient shows someone else or any Unverified Threads … placeholder (account or conversation), they should decline it.

The CLI requires write correlation, selected-account consent, authorization, and quota checks when send is invoked. Treat any rejection as authoritative and do not reroute the send.

threads-messages-cli send --account-id <threads_account_fbid> --thread-fbid <thread_fbid> --text "Hello"
threads-messages-cli send --account-id <threads_account_fbid> --recipient-user-fbids <recipient_fbid> --text "Hello"
threads-messages-cli send --account-id <threads_account_fbid> --thread-fbid <thread_fbid> --media-fbid <visible_threads_post_fbid>
threads-messages-cli send --account-id <threads_account_fbid> --thread-fbid <thread_fbid> --media-fbid <visible_threads_post_fbid> --text "Post context"
threads-messages-cli send --account-id <threads_account_fbid> --thread-fbid <thread_fbid> --text "Reply" --reply-to-message-id '<opaque_message_id>'

--text and --media-fbid are optional individually, but at least one is required. --media-fbid must be a canonical FBID for a visible, published Threads post. --reply-to-message-id is an optional non-empty opaque message ID from an inbox or thread response; never parse it as an FBID. When both media and text are present, the service sends the post share first and the text second. Public stdout describes only the trailing text message when both are present, or the share for a media-only send. In either case it contains exactly the opaque message_id and numeric timestamp_ms. There is no --file option and this CLI does not upload media; it can only share an existing visible, published Threads post by FBID.

Never send until the explicit Hatch confirmation window is approved. A decline or dismissal ends the action. The confirmation makes a best-effort attempt for up to 30 seconds to resolve the sending account, recipient context, and optional media_fbid, but never looks up the opaque reply_to_message_id. Resolved media is rendered as a plain Shared post field, never a link, and any reply is rendered as the literal plain field Reply: Existing message. A lookup timeout, error, malformed response, or missing/unsafe label uses a non-identifying Unverified Threads … placeholder and continues to explicit approval; raw backing account, thread, recipient, media, and reply identifiers are never substituted into the preview. Multi-recipient fallback preserves the exact recipient count and uses placeholders for the whole list if any label is unresolved, so partial lookup cannot understate the audience. A media-only fallback identifies the action as Share Threads post and retains an unverified post placeholder. Supplied message text is rejected only when its trim is empty. Every accepted byte, including leading/trailing whitespace and Unicode, reaches both the structured preview and dispatched request unchanged; platform readback may later show normalized text. Opaque reply_to_message_id values are rejected only when trimming leaves them empty; every accepted value, including one with surrounding spaces, is forwarded byte-for-byte unchanged. If the complete preview cannot fit without truncation, the send fails closed.

No unattended sends

Only send from a conversation where the user approves that exact message. Never send from a scheduled job, cron, watch, background worker, auto-reply, or script, and never set one up that sends Threads messages, even when the user asks for automatic replies. For recurring or automated requests, let the job read and draft, then deliver the drafts to the user to approve one at a time. Never write a job body, memory, or note that claims standing approval to send, and never promise automatic sending. In drafts, do not invent facts about the user or add commitments, promises, or money arrangements the user has not stated; when a reply needs one, leave it out and note what the user needs to decide.

Never use retries for sends. Each separately approved invocation is a new, non-idempotent send. Any failure, including a rate limit, is terminal.

Avoiding duplicate sends

• Run one send at a time and wait for its result before starting the next. Do not fire several sends in parallel; their confirmations can all expire and nothing goes out.
• A send that returned message_id was delivered. Report it as sent.
• After an interruption, cancel, timeout, HTTP 500, or any unclear result, do not send again. Fetch the thread with thread and check whether the message arrived. If it did, report it as sent. If it did not, tell the user it may not have been delivered and send again only after the user asks for a new send and approves it.
• Never loop sends in a script or retry a failed send automatically.

Operating rules

1. This skill is only for the authenticated user's own Threads messages.
2. Reuse a previously fetched Threads account id when you already have it. If you do not, fetch it via threads-cli accounts before using this CLI.
3. Budget API calls against the task at hand. Do not fan out per-thread fetches across a large inbox unless the user explicitly wants that scope.
4. Only paginate when the user actually needs more results. Do not automatically fetch every page.
5. Avoid requests to persistently or frequently poll these commands.
6. Treat U18 enforcement and filtered results as authoritative. Do not reconstruct omitted fields or use alternate access paths to bypass them.
7. Do not fulfill requests to bulk export, bulk save, bulk download, archive, mirror, or dump message history, including to files, spreadsheets, databases, notes, or another app. Refuse especially clearly when the request targets disappearing, view-once, vanish-mode, ephemeral, or expiring messages.
8. For send, require explicit user intent, the exact destination, all exact content, and any exact reply target before invoking the separately approved write.
9. Treat recipient-type and multi-recipient group eligibility failures as authoritative. Do not derive a fallback or reroute a rejected send.
10. Never send Threads messages unattended, including from cron jobs, watches, auto-replies, or scripts. Every send needs the user's approval of that exact message in the conversation.

Output

The CLI prints decoded JSON to stdout.

Read response shape

inbox and thread return the provider's GraphQL shape, not a flat list. There is no top-level threads or messages key, so never read one and conclude the inbox is empty.

• inbox: threads are at data.get_slide_mailbox.threads_by_folder.edges[].node. When threads_by_folder.page_info.has_next_page is true, pass threads_by_folder.page_info.end_cursor as --after for the next page.
• thread: the thread is data.get_slide_thread, which is null when the thread cannot be returned. When its messages.page_info.has_next_page is true, pass messages.page_info.end_cursor as --after for older messages.

Each thread node has thread_fbid, thread_name, thread_type, is_group, folder, timestamp_ms, participants.nodes[] (name, username; no user IDs), messages.edges[].node, and messages.page_info. Each message node has message_id, sender (name, username), sender_fbid (the sender's numeric Threads user ID), content_type, content, and timestamp_ms. Message text is usually content.text_body; other content uses content.text_fragments[].plaintext, content.xma_text_body, content.attachments, or content.videos.

The inbox is empty only when threads_by_folder.edges is an empty array. If the user asks about message requests, also check --folder PENDING; the default INBOX folder does not include them.

Presenting results

Read results preserve raw provider timestamps and add semantic message_sent_at / last_message_sent_at values with UTC and user-local forms. Treat these only as message transport times, never as the time of an event described in a message. send public stdout contains exactly message_id and integer JSON timestamp_ms; the message ID is opaque and must not be parsed as an FBID. When presenting results to the user, focus on meaningful content such as participants, message text, user-local times, links, and media summaries, and avoid exposing raw IDs, cursors, unix timestamps, or implementation details unless the user explicitly needs them for a follow-up command.