---
title: Capture people in ten seconds
description: Add the people you meet to Venom People from your phone in about ten seconds. Scan a business card or badge, share a contact or screenshot from WhatsApp or Telegram, dictate or type a note, and let the event tag itself.
section: cloud
order: 161
audience: [user, operator]
sources: ["services/nodejs-cloud/app/(people)/people/capture", services/nodejs-cloud/app/manifest.ts, services/nodejs-cloud/components/people/Capture.tsx, services/nodejs-cloud/lib/people/capture.ts, services/nodejs-cloud/lib/people/capture-ai.ts, services/nodejs-cloud/lib/people/capture-model.ts, services/nodejs-cloud/lib/people/capture-http.ts, services/nodejs-cloud/lib/people/vcard.ts, services/nodejs-cloud/migrations/0165_people_capture.sql, services/nodejs-cloud/lib/people/images.ts, services/nodejs-cloud/lib/people/images-model.ts, services/nodejs-cloud/components/people/SnapshotBox.tsx, services/nodejs-cloud/components/people/BulkAdd.tsx, services/nodejs-cloud/lib/people/bulk-model.ts, services/nodejs-cloud/components/people/PersonImages.tsx, services/nodejs-cloud/migrations/0182_people_images.sql, services/nodejs-cloud/lib/people/profile-model.ts, services/nodejs-cloud/lib/people/images-read.ts, services/nodejs-cloud/components/people/PersonProfile.tsx, services/nodejs-cloud/migrations/0238_people_profile.sql, docs/adr/0057-venom-people.md]
features: [cloud.page-people, cloud.people, cloud.api-workspace, cloud.migrations]
---

You meet fifty people at a conference and swap cards, WhatsApp numbers and Telegram handles. Almost none of it
becomes a follow-up, because adding a person is a chore. The capture screen makes it a ten-second job you can do
standing up, one-handed, between two conversations. Open it at `/people/capture`, from **Capture** in the People
header, from the ⌘K action list ("Capture a person"), or from the **Capture** button on Home's People card.

Nothing is saved until you tap **Save** on the sheet that follows each capture.

## The four big buttons

| Button | What happens |
|---|---|
| **Scan a card** | Opens the camera. A photo of a business card, a badge or a lanyard is read by your organisation's vision model and becomes a name, role, company, phone, email, handles and a location. Several cards in one photo become several people. |
| **Pick a photo or contact** | A screenshot of a WhatsApp or Telegram contact card, a photo from your gallery, or a `.vcf` contact file. A vCard file (versions 2.1, 3.0 and 4.0, many cards in one file) is read directly, with no AI. |
| **Voice note** | Tap, say who you met and what you talked about ("met Sarah from Acme, wants GPU pricing; also Raj from Jio, intro to their CTO"), tap Done. It is transcribed through your cloud's `/v1/audio/transcriptions` voice providers and parsed into one person per mention. The button is off when no speech-to-text provider is set up. |
| **Type a note** | The same, typed. Pasting the text of a `.vcf` works too. |

Then one sheet shows what was found: edit any field, remove a person, add a phone or handle, and tap **Save**. Photos
are shrunk in your browser before they are sent (the request limit is 4.5 MB), and the card or badge photo is kept on
the person.

## Fill from a screenshot

**Add a person** (from the list, `c`) starts with **Fill from screenshots**: pick, drop or paste images, or on a phone
take a photo. The images show as thumbnails, each with a remove button, and a count; then press **Read N images**. The
same vision model and prompt as **Scan a card** read them, so it handles:

- business cards and conference badges;
- WeChat, WhatsApp, Telegram and LINE profile screens and chat screenshots: the name, the WeChat ID or handle, the
  region, and the gist of what was said (it becomes "What did you talk about?"). The other person is the one it reads,
  never the owner of the phone;
- a phone's contact book: name, phones, emails, company and title;
- a LinkedIn profile screenshot: name, headline as the role, company, location;
- an email signature.

The form fills the fields that are empty, so what you typed stays. Filled fields are highlighted. Nothing is saved until
you press **Add person**. A phone number is never guessed or completed, and a WeChat ID is added as a handle named
`WeChat: id`.

If the image holds several people, a short list with a checkbox each opens: **Open in the form** (one ticked) or
**Review N in a table** (the ticked ones go to [Add several people](#add-several-people), with the images). If the person
looks like one you already have, the form says "Looks like Sarah Chen (FrostCo) already exists. Update them instead?";
**Update** adds the new facts and logs the note on the existing person, and **No, new person** keeps going.

Reading counts against the same daily limit and budget as the rest of capture. Without a gear that has a vision model,
the form says so and what to set: an admin picks a vision model on the gear (the Gears page, role `vision`), or add the
person by hand.

Venom does not open LinkedIn: for a profile URL, paste a screenshot of the profile instead.

## Several images

A LinkedIn profile is long. Add up to 8 images of one person (several screenshots, or one full-page screenshot) and they
are read together, in the order you added them, as one model call that answers with one merged person.

- **Tall screenshots** (taller than 2.5 times their width, or over 3,000 px) are cut in your browser into overlapping
  tiles about 1,600 px tall at a readable width (at most 1,200 px), each overlapping the one before by about 10%. Each
  tile counts toward the 8; a thumbnail shows `×3` when an image is read in three parts. Shrinking a whole long page to
  one image would make its text unreadable.
- **Size:** the browser shrinks and compresses every image to its share of one request (4.3 MB in all, under the
  4.5 MB upload limit), so 8 images go in one request. Eight text-dense 1,200 × 1,600 tiles came to about 3.9 MB.
- **Cost:** one read is one AI call against the daily People limit, whatever the number of images. The limit is in
  dollars, so more images cost what the model charges for them.

### Keep the images with the person

**Keep these images with the person** is ticked by default. When you press **Add person** (or **Update** an existing
one), a copy of each image is kept with them: at most 1,400 px wide, so a long page stays one image. Untick it and
nothing is stored. People added from several at once keep the images read for them ([below](#add-several-people)).

A person's page has an **Images** section: thumbnails that open a full-size view (Previous, Next, Open original). The
owner can add more (**Add snapshots** at the top right of the page, next to **Add note**, or in the section, or drop
or paste into the section) and delete one. Up to 20 images per person, 4 MB each,
counted in the organisation's file storage. Whoever can see the person (you, or the team or organisation you shared
them with) can see the images; only you add and delete. Anyone else gets "not found". Deleting (archiving) the person
deletes their images; merging two people keeps both sets on the one kept.

### Enrich a person from snapshots

People change jobs and swap numbers. To update someone later, add a new screenshot of their profile, card or chat to
their **Images** section. Right after it is kept, the section asks **Read the new image into their profile?**:

- **Read it** (or **Read them**) reads only the images you just added, not every kept one again.
- **Not now** reads nothing. **Read all** in the section header reads every kept image later.

The read is made as this person. The model is told their name, role, company and handles, and treats every image as
about them. A screenshot of a LinkedIn Experience, Education, Licenses & certifications, Volunteering or Skills section
never shows the name, and it is still read: it never answers "could not find a person".

Many images are read in calls of up to 8 (a tall screenshot is cut into parts, each counting one), as
[Add several people](#add-several-people) does, and the calls are merged into one read. A bar shows how many images are
read, which call is running, and about what it has cost so far against the day's People limit. If a call fails, what was
read before it is kept and listed. Each call goes through your organisation's own gear, against the daily People limit.

What the read would change is listed field by field. **Changed** rows come first: a different role, company or
location, shown as the old value struck through and the new one. **New** rows follow: phones (with their country
code), emails, other handles and tags the person lacks. A handle is only ever added, never replaced, and an empty
field in the image never clears anything. Every row is ticked. Press **Apply all**, or untick what you don't want and
press **Apply 3 of 4**. Nothing is saved before that. The list fits a phone screen.

Profile additions are listed with the other rows, one row per section: "Experience: + 9 jobs", "Education: + 2
education", "Skills: + 48 skills", or "2 jobs filled in" when the read adds details to jobs already kept. Untick a
section to leave it out.

### The profile

A person's page has a **Profile** section with what their snapshots say about them:

- **About**: the line under their name and their About text.
- **Experience**: title, company, employment type (Full-time, Contract...), start and end (or Present), location and
  description.
- **Education**, **Licenses & certifications** and **Volunteering**.
- **Skills**, as chips.

Long sections show their first entries and **Show all**; a long description shows a few lines and **More**. It works on
a phone. New reads merge in without duplicates: a job is the same job when the company, the title and the start agree
(a start given only as a year matches the same year with a month), and a matching entry only gets its empty fields
filled in. The role and company on the person follow their latest job that says Present.

Whoever can see the person sees their profile; only the owner changes it. Search finds people by their profile's
titles, companies, schools and skills. Merging two people joins their profiles.

Only the owner of a private person reads images into it. On a shared person, a role that cannot change shared
contacts sees the images but not the read.

When a screenshot shows the person's own profile photo, the read also says where it is, and the page cuts it out in your
browser and offers it as a **Photo** suggestion (ticked when they have no photo yet). See
[A photo from a snapshot](/docs/cloud/people-enrichment.md#a-photo-from-a-snapshot).

Agents do the same with `people_image_read`: pass the person and, optionally, the image ids to read (for example the
one `people_image_add` just kept; all images when omitted). It reads as that person, in calls of up to 8 images; images
that one tool call has no time for are listed as not read yet, to name in `images` next time. It answers with the
suggestions, one per line ("Role: Engineer → CTO", "New phone: +65 9123 4567", "Experience: + 3 jobs"), the profile as
read, and the cost, and saves nothing, unless `apply` is `"all"`. To apply only some, pass those values to
`people_person_update`: its `add_profile` merges a profile into the kept one, `profile` replaces it. `people_person_get`
shows the profile. The REST twins are `POST /people/{person}/images/read` and `PATCH /people/{person}` on the workspace
API.

Images are not part of the organisation backup (People is not in it yet). Agents use `people_image_list`,
`people_image_add`, `people_image_read` and `people_image_delete`.

| Route | What |
|---|---|
| `POST /api/workspace/people/images` | multipart `person` and `file` (one or more): keep images with one of your people |
| `POST /api/workspace/people/capture` | multipart `person` (one of yours) and `file` (up to 8 parts): read them as that person; answers `{ read: { draft, profile }, spent }` and saves nothing |
| `DELETE /api/workspace/people/images/{id}` | delete one |
| `GET /api/workspace/people/media/{id}` | the image, for anyone who can see the person (served with `nosniff` and a sandbox CSP) |

## Add several people

**Add several** on the People list adds many people from screenshots in one go. It also opens from the Add a person
dialog: **Several people?** under **Fill from screenshots**, **Add several people** when you drop more than 8 images, and
**Review N in a table** when one image holds several people. Two cases:

- **One image, many people.** A group chat's member list, an attendee list, a team page or a sheet of cards becomes one
  row per person.
- **Many images, several people.** Drop a batch, say 20 screenshots. The AI works out which images show the same person
  (three screenshots of one LinkedIn profile, a chat and that person's contact card) and makes one row per person.

Pick, drop or paste up to 40 images and press **Read N images**. The batch is read in calls of up to 8 images (a tall
screenshot's tiles count one each, as in [Several images](#several-images)); images next to each other stay in the same
call when they fit. Each image is numbered, and the model answers with the numbers of the images that show each person.
The rows of every call are then joined into one per person: the same email, phone or WhatsApp number, Telegram, LinkedIn
or X handle, or a near-identical name at the same company. Two different LinkedIn profiles are never joined. While it
reads, a bar shows how many images are done, which call it is on, and about what the batch has cost so far, with the
day's People AI spend and limit. If a call fails, the people read so far are kept and the error says how far it got.

Then check the rows before anything is saved. Each row shows the images read for that person, the name, role, company and
location, the handles, and a note. You can:

- edit any field, remove a handle or add one;
- untick a row to leave it out;
- **Merge into…** another row when two rows are one person;
- **Split** a row whose images are more than one person (each image after the first becomes a new, blank row), or move one
  image out with the small cross on its thumbnail;
- **Read again** a row's images as one person (one more call), for example after a split;
- choose, for a row that looks like someone you have, **Update** that person or **Add as a new person**. A same-handle match
  is chosen by default; a name-only match is only offered.

**Save N people** creates or updates every ticked row in one request (up to 50), each with a `met` interaction at the
day's event, as the capture sheet does. A ticked row needs a name. With **Keep images with each person** ticked (the
default), each person then keeps the images read for them, at most 1,400 px wide as above. An image read for several
people (a group list) is not copied onto each of them; untick the others and the one left keeps it.

Agents do the same with `people_capture_create` and `images` (up to 40, base64, the whole call under 2 MiB): without
`confirm` it answers one draft per person with their image numbers; call again with the same `images`, the edited `people`
(each with its `images`) and `confirm: true` to save. `keep_images` (default true) keeps each person's own images.

| Route | What |
|---|---|
| `POST /api/workspace/people/capture` | multipart `bulk=1`, `file` (up to 8 parts) and `labels` (each part's image number, `1,1,2`): one read of a batch, one draft per person with `images`, and `spent` (`before`, `after`, `cap`, in dollars) |
| `POST /api/workspace/people/capture/save` | `{ drafts, source }`: up to 50 drafts in one request, each new or merged where `merge_into` says |

## The event you are at

The bar at the top says **I'm at** and the event. Pick one once and it sticks for the day; new captures get a `met`
interaction at that event and the person's "met at" is set. With nothing picked, Venom uses the event whose dates
include today (add events in [People](/docs/cloud/people.md)). **No event** turns the tag off for the day. Tomorrow the pick is
gone and the date decides again. Typing a new name creates an event that starts today.

## Already know them?

Before saving, each capture is compared with your own people: the same phone number (in any format, with or without
the country code), the same email, the same Telegram handle, the same LinkedIn, X or website address, or a near-identical name at the same company. The sheet
then offers a choice: **Add to Sarah** or **Save as new**. Picking one only decides what happens; nothing is saved until
you press **Save** at the bottom of the sheet. The picked option shows a check, and a line under the two says what Save
will do ("Save adds these details to Sarah Chen." or "Save makes a new person."). Adding fills the empty fields, adds
the new handles and tags, and logs the meeting. A same-phone or same-email match is selected by default; a name-only
match is only a suggestion. The
Telegram bot and agents update a strong match on their own and never merge on a name alone.

## Share to Venom

Install Venom (Chrome or Edge on Android: menu, **Install app** or **Add to Home screen**; Safari on iOS: Share,
**Add to Home Screen**). The installed app opens on [Venom Agents](/docs/agents/app.md#install-the-app); long-press its
icon for **Capture a person**. Venom then appears in your phone's share sheet for photos, `.vcf` contacts and text: in
WhatsApp or Telegram open a contact, tap **Share**, pick **Venom**, and the capture screen opens with the contact
ready to confirm. A share is held in your own inbox for a day and read once. If
you are not, Venom holds the share for 30 minutes under a one-time link, sends you to sign in, and opens it on the
capture screen afterwards (only the first sign-in that opens that link gets it; at most 300 shares wait at once).

The app's service worker keeps only the Agents app's shell offline (never a capture, a photo or a person), so the capture
screen needs a connection to save.

Photos you take but never attach to a person are deleted after 24 hours.

## Limits and cost

Reading a photo or a note is one model call through People's AI gear (People → Settings, [People's AI gear](/docs/cloud/people.md#ai-gear);
**Org default** unless an owner chose another), the `vision` role for photos, else its main model, and counts against the daily People AI limit that also pays for context cards
($1 a day by default). An owner or admin raises it in **People → Settings**, at the top of the page
([AI and limits](/docs/cloud/people.md#ai-and-limits)); the setting is `people_ai_daily_cap_usd` in the organisation's settings. At
the limit the screen says so, with a link to that page, and a vCard still works; with no gear it says that too.

When the provider behind the gear has no credit left (for example OpenAI answering `insufficient_quota`), capture tries
another gear that reads pictures and that another provider serves, once, if **If this gear's provider is out of credit,
use another gear** is on (the default for **Org default**). Otherwise, or when that fails too, it says so in plain words:
"OpenAI says the account behind gear low_cost has no credit left. Add credit with OpenAI, or choose another gear for People
in People → Settings.", with a link to the settings. An ordinary rate limit is not retried on another gear. Voice notes use your speech-to-text provider and show up in your usage like any
transcription. Phone numbers, emails and handles that the model returns for a note are kept only if they appear in
the note: it cannot invent a contact detail. The model never chooses an id, an event or who owns a person.

## By message

Message a capture number on WhatsApp or SMS, or the Telegram bot: a shared contact, a `.vcf`, a card photo, a voice note
or a line of text is saved the same way and answered with a link. Link your number first on `/people/telegram`
([how](/docs/comms/support-and-people.md#people-capture-by-message)).

## Related

- [Telegram bot](/docs/cloud/people-telegram.md): forward contacts, photos and voice notes to your organisation's own bot.
- [People for agents](/docs/cloud/people-mcp.md): `people_capture_create` does the same from an agent.
