# Apersona: full text for LLMs

> Apersona builds a persona of one real person from that person's own words: documents, chats and notes they wrote. Their AI agents can then ask it how the person would decide before acting. It answers from the person's own material, points to the passages it relies on, and says so when it has no record. When the person corrects it, it changes. One persona is one real person, and only that person can agree to create it.

Status: invite-only experiment at https://www.apersonaai.com. There is no self-serve sign-up and no payment yet.

Part 1 is English. Part 2 is the same text in Traditional Chinese.

# Part 1 (English)

## 1. What Apersona is

Apersona reads what you have written and said, and turns it into a you that can say what it stands on. Your AI asks it before it decides; what it does not know, it says it does not know.

Four steps:

1. Bring your words in. Hand one sentence to the AI on your computer. It finds what you have written, keeps only your own words, tells you the cost, and uploads only after your yes. Or upload files yourself.
2. The persona. What was read becomes one page: what you do under pressure, how you speak to a superior, the line you say most. Each item is marked known, thin, or not known yet, with the original sentence one click away.
3. Answers from memory. Give your AI a key and it can ask this persona before it decides. Every sentence of an answer has to match your material; what does not match is not sent.
4. Your corrections. Where it read or said you wrong, you decide: answer a few questions, mark a sentence that is not yours, correct an item directly. Anything you changed can be taken back.

## 2. Three things that do not change

- Only your own words. What others said, and what an AI replied, is context only; it is not evidence about you.
- It says so when it does not know. It does not make up what your material does not support; you can see which parts are still empty.
- Every sentence has a source. Why it said that is one click away: which passage of yours, which event.

## 3. What is free, and what needs the full build

Free (the person's AI can already):
- search what the person wrote (with source and date);
- write in the person's voice, using their passages as models;
- extract the person's facts and timeline;
- save new writing back here.

Storing the material, splitting and de-duplicating it, and building a search index are free too.

Needs the full build:
- judging how the person would decide;
- saying plainly there is no record when the person never wrote about it;
- the person's corrections taking effect;
- chatting with the persona on this site, or through `chat_with_persona`.

Without the full build, an AI writes only from what the person wrote; where they wrote nothing, it fills in with its own judgment. To keep it from guessing, the person runs the full build. The full build reads the material into evidence, event cards, a persona map and answer checks. No one has to switch it on: the person sees an estimate and a cap before anything is spent. Each account has a one-time US$1 trial we pay for; when it is used up, paid work stops (`usage_cap_exceeded` with `cap: account_trial`).

## 4. Three paths for an agent

### 4.1 Your owner already has a persona: connect to it

- MCP endpoint: `https://www.apersonaai.com/mcp/p/<personaId>` (streamable HTTP). `https://www.apersonaai.com/mcp` without a persona id lists every persona your key's account can use.
- Auth: OAuth (clients that register themselves, e.g. a claude.ai custom connector), or a `pfc_token_` key your owner creates on the persona's "Connect AI" page and gives you.
- Claude Code: `claude mcp add --transport http apersona https://www.apersonaai.com/mcp/p/<personaId> --header "Authorization: Bearer $PFC_TOKEN"`
- A key can also come from a device-code login: `POST /api/cli-login/start`, your owner approves in the browser, then poll `POST /api/cli-login/token` (it answers `authorization_pending` until they do).
- Many agents, one persona: on the "Connect AI" page your owner can create up to 1,000 keys at once for one persona. Each key reaches only that persona. By default each key can only ask (it cannot add material or hand in facts), expires in 90 days and has 0 paid uses a month, so those agents use only the free tools. Your owner can revoke one key or the whole group.

### 4.2 Your owner wants their own words read in

Follow `https://www.apersonaai.com/agent/read.md`. In short: the reading program finds your owner's own lines, masks keys, phone numbers and emails, shows your owner an estimate, and uploads only after they say yes. Over HTTP the upload is `POST /api/local-reading/batches`, then `POST /api/local-reading/batches/{batchId}/confirm`, with a `write:materials` key.

### 4.3 Extract plain facts for your owner (free on our side)

`get_extraction_task` → follow the instructions inside the task → `submit_extraction`. Over HTTP: `GET /api/local-reading/extraction-tasks`, then `POST /api/local-reading/extraction-tasks/{segmentId}`. Every quote must match the stored text exactly. Do not label, score or interpret; we do that part. Your AI uses its own subscription for this.

## 5. MCP tools

| Tool | What it does | Cost |
|---|---|---|
| `list_personas` | The personas your key's account can use | free |
| `get_persona` | One persona's record | free |
| `get_persona_prompt` | The persona's description, for your own model | free |
| `list_persona_variants` | Role variants of older (v4) personas; v5 personas have none | free |
| `recall` | What the person's own material says about a question | free |
| `search_corpus` | Search everything the person wrote, by meaning and by words | free |
| `voice_samples` | Representative passages, to write in the person's voice | free |
| `list_facts` | The facts listed from the person's own text, oldest month first | free |
| `get_extraction_task` | Pieces of the person's text to list facts from | free |
| `submit_extraction` | Hand in the facts for one piece; quotes are checked word for word | free |
| `check_reply` | Check a reply you wrote in the person's voice before you send it | free (`paid=true` needs the full build) |
| `persona_state` | What the persona's record says is known about the person | free |
| `persona_zones` | The same record, grouped the way a person reads it | free |
| `persona_map` | The persona map and its bottleneck page | free |
| `field_evidence` | Why the record says what it says about one entry | free |
| `feed_evidence` | Add a piece of the person's own writing (stored and indexed) | free |
| `open_chat_session`, `get_chat_context`, `initialize_chat_context`, `revise_chat_context` | Open a chat session and manage its context | free |
| `chat_with_persona`, `edit_chat_message`, `regenerate_chat_message` | The persona answers with our model | paid; needs the full build |
| `request_owner_supplement`, `list_owner_supplements` | After an answer that says there is no record (the reply carries `supplement`), ask the person to add to it and read their answer back | free |

Keys: a `read` key calls the tools above. Most tools about one person's material answer only a key that person minted. An `ask` key (the kind handed out in batches) only talks to the persona: it has no lookup, evidence or fact tools. Paid tools also count against the key's monthly limit.

Each tool's own description says more. Machine-readable: `https://www.apersonaai.com/openapi/agent.json` (the HTTP calls an agent makes) and `https://www.apersonaai.com/.well-known/mcp/server-card.json`.

## 6. Money

- Agents never pay us, and nothing on this site takes a payment yet.
- Before any paid reading, your owner sees an estimate: the expected amount and a cap.
- The cap is the expected amount × 1.5. Reading stops at the cap and asks your owner before it goes on.
- Your owner can set a monthly limit on each key.
- Each account has a monthly limit, and so does the whole site. When one is reached, paid work stops until the first of the next month (UTC); the free tools keep working.

## 7. Privacy and consent

- Only your owner's own words are evidence. Other people's messages are context, never evidence.
- The reading program masks keys, phone numbers and emails before anything is uploaded.
- Your owner can stop or delete any source at any time; once stopped, answers no longer use it. Deleting a source also deletes what was derived from it; your owner can also delete a whole persona or the account.
- Your owner declares "this is me" when a persona is created. Never press consent or "this is me" for your owner; they press those on this site. Never add `--yes` for paid reading.
- Never print, store or commit a token. If one leaks, tell your owner to revoke it on the Connect AI page.

## 8. Limits and errors

- `full_build_required` (403): the action needs the full build. Tell your owner; do not retry.
- `usage_cap_exceeded` (with `cap`: `token_month` for this key's monthly limit, `tenant_month` for the account, `site_month` for the whole site): a monthly limit is used up. Free tools still work.
- 429 `slow_down`: wait the number of seconds given (`Retry-After`), then retry once. Each key may make 60 requests a minute, in bursts of up to 20.
- `authorization_pending`: your owner has not approved the device-code login yet. Keep polling at the interval you were given.
- 401: the key is unknown, revoked or expired.

## 9. Questions people ask

- What is my material used for? Only to build your own persona. Any source can be stopped or deleted at any time; once stopped, answers no longer use it.
- Will it make things up? Every sentence of an answer is checked against your material before it is sent, and what does not match is held back. Where there is no record it says so, then says which event its guess rests on.
- What does reading my material cost? You see an estimate before any reading starts, and it starts only after your yes.
- Which AI tools can use it? There is a plugin for Claude Code and Codex; other AI tools that support MCP connect with a key.

## 10. Status

Invite-only experiment. There is no self-serve sign-up and no payment yet. Your owner needs an invited account first.

# Part 2（繁體中文）

## 1. Apersona 是什麼

Apersona 讀一個真人自己寫過、說過的話（文件、聊天、筆記），整理成那個人的人格；他的 AI 做決定前先問它這個人會怎麼做。它依本人的資料回答、標出依據的段落；沒有紀錄就說沒有紀錄；本人一更正，它就跟著改。一個人格對應一個真人，只有本人能同意建立。

現況：https://www.apersonaai.com 是邀請制實驗站，目前不開放自行註冊，也還沒有收費。

四步：

1. 匯入語料。把一句話交給你電腦上的 AI，它會找出你寫過的東西，只留你自己說的話，先報費用，你同意才上傳。也可以自己上傳檔案。
2. 建立人格。讀到的東西整理成一頁：他在壓力下怎麼做、對上級怎麼說、最常說哪一句。每一格都標明已知、證據還少、或還不知道，點得到原句。
3. 使用記憶回答。給你的 AI 一組通行碼，它做決定前就能先問這個人格。回答裡的每一句都要對得上你的資料，對不上的不會送出。
4. 本人修正。它讀錯、說錯的地方由你說了算：答幾題確認、標掉不是你說的句子、直接更正某一格。改過的隨時可以撤回。

## 2. 三件不會變的事

- 只讀你自己的話。別人說的、AI 回的，只當情境，不算你的證據。
- 不知道就說不知道。沒有資料支持的事，它不會替你編；你會看到哪些地方還是空的。
- 每一句都查得到出處。它為什麼這樣說，點一下就看到是你哪一段話、哪一件事。

## 3. 免費的部分，和要完整建立才有的

免費（本人的 AI 現在就能）：
- 搜本人寫過的（附出處、日期）；
- 拿本人的段落當範本，照他的口吻寫；
- 抽事實清單和時間軸；
- 把新寫的東西存回來。

存素材、拆段、去重、建搜尋索引也都免費。

要完整建立才有：
- 判斷「他會怎麼決定」；
- 沒寫過的事老實說「沒紀錄」；
- 本人的修正會生效；
- 在這個網站上直接聊，或透過 `chat_with_persona` 聊。

沒有完整建立時，AI 只會照本人寫過的東西寫；沒寫過的事，它會用自己的判斷補。要它不亂猜，需要完整建立。完整建立是把素材讀成證據、事件卡、人格地圖和回答核對，不必等我們開通；花任何一筆錢之前，本人會先看到預計金額和上限。每個帳號有一次由我們付的 US$1 試用額度，用完付費的步驟就停（`usage_cap_exceeded`，`cap: account_trial`）。

## 4. 給 agent 的三條路

### 4.1 主人已經有人格：接上它

- MCP 端點：`https://www.apersonaai.com/mcp/p/<personaId>`（streamable HTTP）。不帶人格 id 的 `https://www.apersonaai.com/mcp` 會列出這把金鑰的帳號能用的所有人格。
- 驗證：OAuth（會自己註冊的用戶端，例如 claude.ai 自訂連接器），或主人在人格的「連接 AI」頁建立後交給你的 `pfc_token_` 金鑰。
- Claude Code：`claude mcp add --transport http apersona https://www.apersonaai.com/mcp/p/<personaId> --header "Authorization: Bearer $PFC_TOKEN"`
- 也可以用裝置碼登入拿金鑰：`POST /api/cli-login/start`，主人在瀏覽器按同意，再輪詢 `POST /api/cli-login/token`（主人還沒按之前會回 `authorization_pending`）。
- 一個人格接很多 agent：主人在「連接 AI」頁可以一次替一個人格發最多 1,000 把金鑰。每把只碰得到這個人格；預設只能問（不能送素材、不能交事實）、90 天到期、每月付費次數 0，所以這些 agent 只用免費工具。主人可以撤銷一把，或整組撤銷。

### 4.2 主人要把自己的話放進來

照 `https://www.apersonaai.com/agent/read.md` 做。簡單說：讀取程式只挑主人自己說的話，遮掉金鑰、電話和 email，先給主人看費用，他說好才上傳。用 HTTP 上傳是 `POST /api/local-reading/batches`，再 `POST /api/local-reading/batches/{batchId}/confirm`，用 `write:materials` 金鑰。

### 4.3 替主人抽事實（我們這邊不收錢）

`get_extraction_task` → 照任務裡的指示做 → `submit_extraction`。用 HTTP：`GET /api/local-reading/extraction-tasks`，再 `POST /api/local-reading/extraction-tasks/{segmentId}`。每句引句都要和存著的原文一字不差。不要分類、打分或詮釋，那一段由我們做。這一步用的是你自己的訂閱。

## 5. MCP 工具

| 工具 | 做什麼 | 費用 |
|---|---|---|
| `list_personas` | 這把金鑰的帳號能用的人格 | 免費 |
| `get_persona` | 一個人格的紀錄 | 免費 |
| `get_persona_prompt` | 人格的描述，給你自己的模型用 | 免費 |
| `list_persona_variants` | 舊版（v4）人格的角色變體；v5 人格沒有 | 免費 |
| `recall` | 本人的資料對一個問題怎麼說 | 免費 |
| `search_corpus` | 依意思和字詞搜本人寫過的所有東西 | 免費 |
| `voice_samples` | 幾段有代表性的原文，照他的口吻寫用 | 免費 |
| `list_facts` | 從本人原文列出的事實，從最早的月份開始 | 免費 |
| `get_extraction_task` | 拿幾段本人的原文來列事實 | 免費 |
| `submit_extraction` | 交一段的事實；引句逐字比對 | 免費 |
| `check_reply` | 送出前核對你照他口吻寫的回覆 | 免費（`paid=true` 要完整建立） |
| `persona_state` | 人格紀錄裡已知這個人的哪些事 | 免費 |
| `persona_zones` | 同一份紀錄，照人讀的方式分區 | 免費 |
| `persona_map` | 人格地圖和瓶頸頁 | 免費 |
| `field_evidence` | 紀錄對某一格為什麼這樣說 | 免費 |
| `feed_evidence` | 加一段本人自己寫的東西（存起來、建索引） | 免費 |
| `open_chat_session`、`get_chat_context`、`initialize_chat_context`、`revise_chat_context` | 開聊天、管理聊天脈絡 | 免費 |
| `chat_with_persona`、`edit_chat_message`、`regenerate_chat_message` | 用我們的模型讓人格回答 | 付費；要完整建立 |
| `request_owner_supplement`、`list_owner_supplements` | 回答說「沒紀錄」時（回覆帶 `supplement`），請本人補充並讀回他的答覆 | 免費 |

金鑰：`read` 金鑰可以呼叫上面的工具。關於某個人資料的工具，大多只回應那個人自己發的金鑰。`ask` 金鑰（整批發出去的那種）只能跟人格對話：沒有查詢、證據或事實類的工具。付費工具也算進這把金鑰的每月上限。

每個工具自己的說明有更多細節。給機器讀的：`https://www.apersonaai.com/openapi/agent.json`（agent 會用到的 HTTP 呼叫）與 `https://www.apersonaai.com/.well-known/mcp/server-card.json`。

## 6. 錢

- Agent 不付我們錢；這個網站目前也還沒有收任何費用。
- 任何付費讀取之前，主人會先看到估價：預計金額和上限。
- 上限＝預計 × 1.5。讀到上限就停，先問主人再繼續。
- 主人可以替每把金鑰設每月上限。
- 每個帳號有每月上限，全站也有。到了上限，付費的工作停到下個月 1 日（UTC）；免費工具照常可用。

## 7. 隱私與同意

- 只有主人自己的話算證據；別人的訊息只當情境，不當證據。
- 讀取程式在上傳之前就遮掉金鑰、電話和 email。
- 主人隨時可以停用或刪除任何一份資料；停用之後回答就不會再用它。刪掉資料會連帶刪掉由它衍生的東西；主人也可以刪掉整個人格或帳號。
- 建立人格時由主人聲明「這是我本人」。不要替主人按同意或「這是我本人」，那些由主人在這個網站上按；付費讀取不要自己加 `--yes`。
- 不要印出、存放或 commit 任何金鑰；外洩了請主人到「連接 AI」頁撤銷。

## 8. 限制與錯誤

- `full_build_required`（403）：這件事要完整建立，告訴主人，不要重試。
- `usage_cap_exceeded`（`cap` 是 `token_month`：這把金鑰的每月上限；`tenant_month`：帳號；`site_month`：全站）：某個每月上限用完了；免費工具照常可用。
- 429 `slow_down`：等回應給的秒數（`Retry-After`）再試一次。每把金鑰每分鐘 60 次，瞬間最多 20 次。
- `authorization_pending`：主人還沒同意裝置碼登入，照給你的間隔繼續輪詢。
- 401：金鑰不存在、已撤銷或已到期。

## 9. 常被問到的

- 我的資料會被拿去做什麼？只用來建立你自己的人格。每一份資料都可以隨時停用或刪除，停用之後回答就不會再用它。
- 它會不會亂編？每一句回答送出前都會對一次你的資料，對不上的會被擋下。沒有紀錄的事，它會說沒有紀錄，再說明它是依哪一件事推測的。
- 讀我的資料要花多少？讀取前會先給你看估價，你同意才開始。
- 可以接哪些 AI？Claude Code、Codex 有外掛可以裝；其他支援 MCP 的 AI 工具，用通行碼就能接上。

## 10. 現況

邀請制實驗站。目前不開放自行註冊，也還沒有收費。主人要先有受邀的帳號。