For people who run many X accounts
AI personas: posts that sound like each X account
One persona per X account: what it talks about, how it writes, what it avoids, its emoji and hashtag habits. Flottra drafts it from a topic and a brief with your own Anthropic key and you edit it. The posting routine, AI-written scheduled posts and AI agents through the MCP server then write in that voice.
Updated
What it does
One persona per account, so twenty accounts do not sound like one bot
A persona is the standing description of how one X account talks. It is not a tweet generator box; it is a setting on the account that everything writing for that account reads first.
Most of what ranks for "AI Twitter bot" is a tutorial for building one account's bot: a script, a model key, a cron job. If you run many X accounts the problem is different. One model with one prompt writes the same way for every account, and twenty accounts that open posts the same way, close them the same way and reach for the same three emoji are the easiest cluster on X to spot.
A persona in Flottra is the fix at the account level. Each X account carries its own: what it talks about, how it writes, the words and moves it never uses, how long its posts run, how often it uses emoji and hashtags, and a handful of example posts that set the cadence. Anything that writes for that account reads the persona first, so the DeFi account stays terse and lowercase while the NFT account writes in full sentences.
You do not have to write one by hand. Give the panel a topic and a brief and it drafts the persona with your own Anthropic key; you read it, change what is wrong and save it. Nothing is stored until you accept the draft.
The three cards beside this text show the result: one product, three accounts, three voices. The sample lines are illustrations, not posts from real customers.
bridged at 3am to save forty dollars. bridge took six hours. saved nothing
- voice
- lowercase, terse, complains about tooling
- emoji
- none
- hashtags
- none
- length
- short
WE ARE SO BACK. do not ask what changed
- voice
- loud, caps for emphasis, one line at a time
- emoji
- some
- hashtags
- rare
- length
- micro
Floor is down 12 percent and my bid is still the top one. Most of this market is people waiting for someone else to blink first.
- voice
- full sentences, patient, a collector explaining himself
- emoji
- rare
- hashtags
- none
- length
- medium
The fields
What a persona holds, and what each field changes
The fields of the persona editor. This is the list from the code, not a brochure version of it.
A persona is a small JSON document saved against one account. The panel shows it as a form; an agent reads and writes the same document through the MCP server. Every field below goes into the instructions the text generator receives, so a change here changes the next draft.
- Bio
- One sentence on who the account is. The generator reads it as identity, and together with the voice it sets the default tone: a bio that says professional or analyst is drafted more formally, one that says snarky or funny gets a looser draft.
- Voice
- Two or three sentences on how the account writes: sentence length, punctuation, how much it uses capitals, how warm or dry it runs. This is the field that does most of the work.
- Interests
- Specific topics, eight to ten in a drafted persona. The generator treats them as what the account cares about, and a draft on a topic outside the list reads as off, which is the point.
- Quirks
- Recurring phrases the account sometimes reaches for. The generator is told to use one in roughly one post in five and never every time, so an account does not turn into a catchphrase.
- Avoid list
- Words, tones and moves the account never uses: corporate speak, a moral at the end, a rhetorical question to fish for replies. The generator is told these are never allowed.
- Emoji usage
- none, rare, some or frequent. Each maps to a rate: none means never, rare means about one post in ten carries a single emoji, some means about half of posts carry one or two, frequent means most posts carry one to three.
- Hashtag usage
- The same four levels for hashtags, with none as the default. No level produces a wall of hashtags at the end of a post.
- Length bias
- Weights for four buckets: micro, short, medium and long. Each draft rolls a bucket by weight, so an account weighted towards micro posts a two-word line one day and a sentence the next, and the mix looks like a person rather than a setting.
- Example posts
- Real past posts. The generator mirrors their rhythm, punctuation and capitalisation, not their topics. Leave them out and it builds the cadence from the voice description alone.
- Search terms
- Short words a real person in this niche would put in a bio. They feed target discovery: when a follow routine is on, the app searches for people with these terms to find accounts worth following.
Three smaller fields sit beside those: a short id used in logs, a gender lean for the implied human, which the drafter picks unless you set it, and a list of subreddits that nothing on the hosted beta reads yet. The editor has no image field.
How it works
Describe the account, let the panel draft the persona, edit and save
Three steps on the Personas page. A few minutes per account, and the draft costs one call to the model on your own key.
- 01
Describe the account
Open Personas, pick the account and press Generate persona. Give it a topic or niche and, if you want control, a brief: who this person is, in a sentence or two. Tone, gender lean, interest seeds and extra notes are optional; the brief wins over all of them. - 02
Let the panel draft it with your key
The panel calls Anthropic’s API with the key you saved under AI provider keys in Settings. The draft is written in the language of the account’s proxy country, so a Turkish proxy gets a Turkish persona, and it comes back as a filled-in editor. - 03
Edit and save
Read the draft. Cut a quirk that sounds off, tighten the avoid list, paste real past posts as examples. Nothing is stored until you press Accept & save persona, and you can reopen the editor later.
The draft is written by Claude Sonnet 4.6, named in the code as claude-sonnet-4-6, with the key from your own Anthropic account. Flottra does not resell model calls and does not bundle a model into the plan. Every call is logged in your workspace's AI request log with the model and the token counts, so you can see what a draft cost you.
The language rule is worth knowing before you assign proxies. The drafter reads the country of the proxy on the account and picks the language from it: a German proxy gives a German persona, a US or UK proxy gives English, and no proxy gives English. If an account should write in a language its proxy does not imply, say so in the brief; the brief is read literally and wins.
You can also skip the drafter and fill the form yourself, or have an agent copy one of eight built-in templates into an account through the MCP server. The result is the same document either way.
- 1
- persona per account
- 4
- length buckets: micro, short, medium, long
- 1
- API key, yours, stored encrypted
- 3
- drafts per post; the least repetitive one is kept
Agents
Who writes from a persona: the routine, scheduled posts and agents
Three things read a persona before they write: the posting routine, the Let AI write it mode of a scheduled post, and an AI agent connected through the MCP server.
The posting routine and scheduled posts are the product writing by itself. The text is written on Flottra's server with the AI key you saved in Settings, from the persona of the account that will send it, and the desktop app on your computer posts it; the app itself never calls a language model. The scheduling page has the form and the routine.
The routine adds a layer under the persona. Each account gets a set of writing habits drawn once and kept: how it uses capitals, which post shapes it favours, its kind of humour. For every routine post three drafts are written, and the one least like what the account and its sibling accounts posted recently is kept.
Flottra ships an MCP server with 108 tools, and five of them are about personas: list, library, get, set and preview. An agent such as Claude Code connected with a workspace key can read every account's persona, set or replace one, and copy a built-in template into an account it is onboarding. Every one of those writes is audited against the key that made it.
The tool that matters on this page is personas_preview. It is a real dry run through the same text generator the product uses: it takes the account's persona, an optional topic or a post to reply to, and the last eight replies the account sent so the draft does not repeat itself, and it returns the text without posting anything. It costs one model call on the workspace's configured provider key and fails cleanly if the workspace has none; it is logged in the AI request log, and it needs a key with write scope for that reason.
From there the agent has text in the account's voice, and it can schedule it as a reply or as a post: the desktop app on your computer opens the account's own browser through its own proxy, posts it, and the confirmed URL lands in the task result. The auto reply page covers replies, including the engagement autopilot that writes them in the persona's voice.
- Passwords, cookies and API keys are never returned by any tool, the persona tools included.
- A persona is per account, so an agent running twenty accounts reads twenty documents and writes twenty ways.
Pacing
A good voice does not skip the queue
Whatever an agent writes goes through the same gates as a task you typed by hand.
Two checks sit between a drafted reply and the account's browser, and neither cares how good the text is. The first is the warming gate: an account's age puts it in a phase, and the phase says which actions are allowed and how many a day. A reply is not on the list until phase 3, which an account reaches after seven days, and the phase 3 cap is 12 actions in 24 hours counting what is already queued. The second is the hourly cap per account, enforced when scheduling through the MCP server: 5 replies an hour by default, 3 posts, 12 likes, 6 follows. A post is not allowed until the ramp is over, 14 days in the default ramp, and the posting routine waits for that too.
So an agent that drafts thirty sharp replies for a three-day-old account gets thirty refusals, each with the reason. That is deliberate. The persona makes an account sound like a person; the gates make it act like one. The caps are per account, so twenty accounts get twenty budgets, and no single account can be made to talk faster than its phase allows.
Beta status
What runs today, and what does not yet
So a tester knows exactly what to expect. Both lists are true to the code on 17 November 2026.
Today, on the hosted beta
- A persona per account: draft it from a topic and a brief with your own Anthropic key, edit it, save it, reopen it.
- The posting routine and the Let AI write it mode of a scheduled post write from it, with your own AI key, and the desktop app posts the text.
- An AI agent through the MCP server reads it, sets it, previews a post or a reply in that voice, and schedules the text for the desktop app to post.
- The engagement autopilot, set up through the MCP server, writes replies in the persona's voice.
- Every draft and preview logged in your workspace's AI request log with the model and the token counts.
- The warming gate and the hourly caps on everything the agent schedules.
Not on the hosted beta yet
- Replies to mentions and notifications in the persona's voice.
- A panel page for the engagement autopilot. Today it is set through the MCP server.
- Images or a photo style in the persona editor.
The rules
What X says about automation, quoted
Read this before you let anything write for an account. The rules apply whether the text is good or not.
You "may not follow or unfollow X accounts in a bulk, aggressive, or indiscriminate manner", you "may not like posts or hide replies in an automated manner", unsolicited automated direct messages are not permitted, and scripting the X website "may result in the permanent suspension of your account".
X also publishes a page on automated account labels, which says its automation rules require automated accounts to display the label and remain connected to a human-run account. The reply bots guide quotes the line that matters most here: the deployment or operation of any AI reply bot requires prior written and explicit approval from X. A persona does not come with that approval. Read the pages yourself before you decide; the rules were updated in April 2026 and will change again.
Flottra does not claim compliance with those rules. It paces everything and refuses bursts, which reduces the rate at which accounts are locked or suspended; it does not make it zero, and nobody can. The automation guide is the long version of what browser automation can and cannot reach in 2026, and the reply bots guide covers why identical text across accounts is what gets caught, which is the pattern a persona per account exists to break.
Questions
Questions people ask about this
- Does Twitter allow bot accounts?
- X allows automation inside its automation rules, and the rules are specific: no bulk or aggressive following, no automated likes, no unsolicited automated replies or direct messages, and scripting the X website may result in permanent suspension. X’s page on automated account labels says the rules require such accounts to display the label and stay connected to a human-run account. Read the current pages yourself; they change. Flottra does not claim to comply and does not promise an account will not be actioned.
- How do I create a Twitter bot account?
- Not with Flottra. It does not create X accounts, buy them or sign them up for you. You bring accounts you already own, add them to the panel with a password or an auth token cookie, give each one its own proxy, and pair the desktop app that runs their browsers on your computer. What Flottra adds on top is the persona per account, the pacing and the MCP server. Creating the account itself is a separate job you do yourself.
- Which AI model writes the personas?
- The persona draft is written by Claude Sonnet 4.6; the model id in the code is claude-sonnet-4-6. The panel calls Anthropic’s API directly with the key you saved in Settings under AI provider keys, so the call is billed to your own Anthropic account, not to Flottra. Each call is logged in your workspace’s AI request log with the model and the token counts. There is no bundled model, and no persona drafting without your key.
- Do you see my API key?
- It is stored encrypted at rest with AES-256-GCM, the same way X passwords, cookies and proxy passwords are stored. The key that decrypts them is a server key held by Flottra, not a key only you hold, so the host can decrypt a stored key when the panel needs it for a call. That is shared-key encryption, not end to end, and the site says so plainly. The MCP server never returns a stored key to any tool; settings report presence only.
- Does Flottra post in the persona’s voice by itself?
- Yes. The posting routine and the Let AI write it mode of the Schedule a post form both write from the account’s persona, on Flottra’s server, with the AI key you saved in Settings, and hand the text to the desktop app, which posts it from the account’s own browser. The app itself never calls a language model. Replies in the persona’s voice come from the engagement autopilot, which is set up through the MCP server, and an AI agent can draft in the same voice with the preview tool.
- Can twenty accounts share one persona?
- A persona is saved per account: one document, one account. You can put the same document on twenty accounts, through the editor or with the MCP set tool and its eight built-in templates, but the product is built for the opposite. Twenty accounts with one persona write the same way, and identical style across accounts is the pattern that gets clusters caught. The bulk pools on the Identity page exist for profile text, bios and display names; personas stay per account.
- Is this an AI tweet generator?
- No. There is no free box where you type a topic and copy a tweet. A persona is a per-account setting inside a workspace that also holds the account’s proxy, browser profile, warming phase and task queue, and it is read by whatever writes for that account: the posting routine, an AI-written scheduled post, the engagement autopilot’s replies, or an AI agent through the MCP server. If you want one tweet, a generator is cheaper; if you run many accounts, this is what keeps them from sounding alike.