On by default, and never posted
Community answers are on by default for every workspace, including existing ones that never touched the setting (community_answers_enabled). If someone switches the setting off, it stays off until someone switches it back on.
You turn them on and off yourself, in the app under Settings › Autonomy (Draft community replies), or with community_answers_enabled on PATCH /v1/settings/agents or the MCP tool update_settings. Off stops new drafts: a draft request is refused with a message that says where to switch it back on. Drafts already in Work stay there.
Switching it off needs no more than the route itself asks for. Switching it on from an API key needs the configure scope. An OAuth connection does not carry configure, so turn it on in the app. Neither direction is a raise of what the agent may do on its own, because a draft is never posted for you; that is why a key may switch it on at all, when it may not switch publishing or sending on.
Drafts are never posted. The agent writes the reply and leaves it in Work for you to read, edit, copy and post from your own account. Attensira never posts, comments or votes on a community: no tool, route or background job can, and there is no setting that changes that, in any approval mode.
If you run your own agent on your own accounts, it can read the drafts over MCP or the REST API, and the thread list with your API key (see Over MCP and the API). What it does with them on your accounts is your action, not ours.
How threads are found
Threads are picked from the deciding sources: the pages AI answers cite for your tracked prompts. Threads on Reddit, Hacker News, Stack Overflow and the Stack Exchange sites, and Quora qualify. Each thread in the list comes with:- Its link, so you can open it where it lives.
- Why it matters: which engine cited it, answering which of your prompts, and how much of that answer’s citations it carried, for example ChatGPT cited it 4 times when answering “best crm for startups” (25% of that answer’s citations). Every answer that cites it is listed with it.
- Its draft, once one has been written.
find_community_threads tool does this lookup. It is free and read-only: it reads only data Attensira already holds, never the community itself. See The agent’s tools.
Opening a thread drafts a reply
To find the list in the app, search for Community threads in the command palette. It is not in the sidebar. Each thread’s button reads Draft a reply, or Open draft once it has one. Draft a reply writes a draft for the thread: the agent finds the AI answer that cites the thread, reads the thread where it is allowed to, and files the draft in Work. Opening the same thread again on the same day hands back the same draft or the same run; it does not start a second one. Sometimes the run decides there is no genuinely useful reply to make. The thread then shows no draft, with See why, a link to the run that says why, and opening it again before midnight UTC does not start another run; after midnight UTC it drafts afresh. If a run fails or is cancelled, opening the thread again tries once more, up to three runs per thread per day (UTC). The thread shows failed only once all three of the day’s runs have failed: “This draft didn’t finish after three tries today. Try again after midnight UTC.” There is no retry button then, because trying again the same day gets the same answer; open the thread again the next UTC day. See why links the run. When the list or a draft cannot load, the screen shows the reason. A timeout, too many requests at once, or running out of credits still offers Retry, because the same request can succeed later. Only a real refusal hides it, because the same request would get the same answer: a bad request, a thread that is not found, a workspace not enrolled in off-site work, or community answers turned off.Reading threads
Attensira reads a community thread only through a route the site allows, and never logged in to anyone’s account. Reddit is read only through Reddit’s official Data API, as an application registered with Reddit under its Data API terms. Every request is authenticated as that application, never as you or any other Reddit user, and nothing reaches Reddit any other way: the web reader refuses every reddit.com address. The API is used to read, and only to read.- Reddit’s published rate limit is kept. Reads stay under the limit Reddit publishes for an application, and follow what Reddit reports on every response. When the limit is reached, the agent is told to try again shortly or to move on; a read is never retried in a loop.
- When Reddit’s API is not available to Attensira, a Reddit thread is not read, and the agent says so and gives the link. The thread list keeps working, because it is built from the AI answers Attensira already reads, not from Reddit.
- A thread Reddit no longer serves (deleted, removed or made private) is not read, and gets no draft.
- robots.txt is honoured. Before reading a thread, Attensira reads the site’s robots.txt and follows it. If robots.txt does not allow the read, the thread is not read, and it is never handed to another reader to get around that.
- Rate-limited per site. At most one read every few seconds per site from each Attensira server, whichever workspace asks. When a site is busy, the agent is told to try again shortly or to move on.
- Never logged in. Reads carry no login, no cookie and no password. The robots.txt check never follows a redirect to another site; the page read itself follows redirects as a browser does.
What Attensira keeps from Reddit
The agent reads the whole thread while it drafts, but Attensira keeps only the following. The record of the agent’s run keeps the thread’s link, not its text.- The thread’s id and link.
- What it worked out about the thread: its score and number of comments, whether it still takes replies, and whether the subreddit has a self-promotion rule.
- Short quotes: the title and up to three top comments, each cut to 200 characters, each tied to the post or comment it came from, with no username kept.
The rules every draft follows
The tool that files a draft refuses one that breaks these, whatever the agent intended:- Disclosure. A reply that names your product must include a plain disclosure line, such as “Disclosure: I work on Acme.” An undisclosed mention is refused.
- One link, last. At most one link to you, alone in the last paragraph, after the reply has answered the question. A link before the answer, or a second link, is refused.
- Self-promotion rules. When the subreddit has a self-promotion rule, a note on the draft says so.
- Account floor. On Reddit, the reply is meant for a real person’s account at least 90 days old with at least 100 karma. When the account’s age and karma are known and below that floor, nothing is proposed.
- No posting, ever. A request to post, submit, comment or vote is refused and nothing is written.
- Evidence first. A draft is made only for a thread an answer actually cites, and it carries that answer’s own words, the engine and the date it was read.
- Rules beat the draft. A subreddit rule that bans promotion or links removes the mention, not the rule. If the rules or the thread could not be read, the draft says so and you read them before posting.
- One reply per thread. Never a second reply, a reply for a colleague to agree with, or a planted question.
- Nothing to add is an answer. If an existing reply already says it, no draft is made.
What the threads tell us
A thread that decides answers also shows what buyers ask, what puts them off, and what they believe about you. The agent records these as insights. Each one comes from a thread your AI answers cited in the last 28 days and carries that thread’ssource_url. Each is filed for you in one of two ways, and never as a reply:
Nothing changes on your site or in your Brand Book until you approve the row. Recording the same insight twice files nothing new.
No insight is recorded from a Reddit thread today: Reddit is read only through its licensed Data API, which is not connected. Insights come from the other communities, such as Hacker News, Stack Overflow, Stack Exchange and Quora.
Recording insights is separate from drafting replies. It works whether community answers are on or off, because it never writes anything for a thread.
Over MCP and the API
The thread list and the drafts are read-only on every door. The one write is the on/off setting:
There is no verb that posts, and none will be added. To have a draft written over MCP or the API, ask for one with
ask.