Describe
Sign inJoin the waitlist

Describe what should happen. A browser does it.

Write the task in plain words. Describe reads it into a blueprint you can check, then an agent carries it out in a real browser you can watch and take over, once or on a schedule.

Watch the demo

We are letting people in from the list in order. Free is 20 runs a month and every model. No card.

Every weekday at 8am, read my unread Gmail from the last 24 hours, group it by what needs a reply and what is waiting on someone else, then post the summary to #ops in Slack.

✦ Claude Sonnet 5.5
  1. ◷Weekdays at 8:00am0 8 * * 1-5

    “Every weekday at 8am”

  2. MGmail

    “read my unread Gmail from the last 24 hours”

  3. ✦Claude Sonnet 5.5

    “group it by what needs a reply and what is waiting on someone else”

  4. SSlack

    “post the summary to #ops in Slack”

You see the plan before anything runs.

As you type, the description is drawn as a diagram: when it runs, then one node per step, each quoting the words it was read from. The same drawing is in the thread when the agent proposes it, on the automation's page once it is running, and in the API as JSON. If the drawing is wrong, the words are wrong, and you fix the words.

◷

The trigger is a time or an event

“Every weekday at 8am” becomes a schedule. “When a new contact is created in HubSpot” becomes an event. A description with neither is a one-off task.

M

Each step names its app

A step that reads mail is attributed to Gmail, a step that posts is attributed to Slack, and a step that only thinks is attributed to the model you picked.

✦

Any model, per automation

The cheap model sorts your inbox. The deep one reconciles the books. Credits are charged by what each run actually used.

A real browser, not a screenshot of one.

Every agent works in its own Chromium with its own profile, on its own display. You watch it live. When a site asks for a login or a file, you take the controls; the file comes from your machine, never from ours. Hand the controls back and the agent carries on.

  • Logins stay in that browser's profile, private to you and to no one else in the workspace.
  • The browser cannot reach the node's disk, DevTools, extensions or printing. Policy, not politeness.
  • Uploads come from your device through the live view, so the agent never holds your files.
A 54-second run, unedited: “go to Redfin, search homes for sale in 21029, sort by newest, give me the five newest as a table.” The agent explains each step on the left while the browser does it on the right.

Things people describe.

Four descriptions from real workspaces, with the drawing each one became. Click any to open it in the app with the words already typed.

Every day at 7am check the pricing pages for Tasklet, Lindy and Zapier. If a plan price or a plan limit changed since yesterday, email me the diff with a one-line take on what it means.

  1. ◷Daily at 7:00am
  2. wwwThree pricing pages
  3. ✦Diff since yesterday
  4. MEmail me

When a new contact is created in HubSpot, look them up, write a two-sentence note on who they are and whether they fit our ICP, add it to the contact, and ping #sales in Slack if they look like a fit.

  1. HNew contact
  2. wwwLook them up
  3. HAdd the note
  4. S#sales

Every Friday at 4pm, pull this week's paid invoices from Stripe and the matching rows in the Revenue sheet; flag anything that does not match and put the list in the sheet's Exceptions tab.

  1. ◷Fridays at 4:00pm
  2. SPaid invoices
  3. SRevenue sheet
  4. SExceptions tab

Every weekday at 8am, read my unread Gmail from the last 24 hours, group it by what needs a reply, what is waiting on someone else, and FYI, then post the summary to #ops in Slack.

  1. ◷Weekdays at 8:00am
  2. MUnread mail
  3. ✦Group it
  4. S#ops

Every model on the picker.

Pick a model per automation, or per message. Cost is marked one to four so the choice is visible before the run, and the run's charge is what the model used, never a flat fee.

Also on the picker: GPT-6 mini, o5, Grok 5, Mistral Large 3.

  • Claude Fable 5.1AnthropicDeepest reasoning; long, careful tasks1M
  • Claude Opus 5.5AnthropicStrong all-round; a good default for agents1M
  • Claude Sonnet 5.5AnthropicFast and capable; most automations500k
  • Claude Haiku 5.5AnthropicCheapest; sorting, routing, extraction200k
  • GPT-6OpenAIBroad knowledge; strong at writing1M
  • Gemini 3.5 ProGoogleLong documents; video and images2M
  • Gemini 3.5 FlashGoogleVery fast; high-volume automations1M
  • Llama 5 405BMetaOpen weights; runs in your own cloud256k

Connected to the apps your business already runs on.

Connect once, share it with the workspace or keep it private. For everything else there is the browser: any web page, any HTTP API, any MCP server.

  • MGmail
  • 31Google Calendar
  • DGoogle Drive
  • SGoogle Sheets
  • DGoogle Docs
  • SSlack
  • NNotion
  • HHubSpot
  • QQuickBooks
  • GGitHub
  • AAirtable
  • LLinear
  • SStripe
  • SSalesforce
  • ZZendesk
  • JJira
  • AAttio
  • AAsana
  • IIntercom
  • SShopify
  • OOutlook
  • TMicrosoft Teams
  • CCalendly
  • XXero
  • DDocuSign
  • PPipedrive
  • wwwWeb pages
  • { }Any HTTP API
  • mcpMCP server

The same reading, as JSON.

What the app draws is what the API returns: a trigger, the steps, the connections, the model. Read a description, check the plan, then confirm it as an automation. Webhooks tell you when a run ends and what it cost.

Keys, logs and webhooks are in the app

# read a description into a plan
curl https://api.describe.ai/v1/describe \
  -H "Authorization: Bearer dsc_live_…" \
  -d '{ "text": "Every weekday at 8am, read my unread Gmail
            and post a summary to #ops in Slack" }'
{
  "id": "rd_7f3a9c",
  "kind": "automation",
  "trigger": { "kind": "schedule", "label": "Weekdays at 8:00am", "cron": "0 8 * * 1-5" },
  "steps": [
    { "connection": "gmail", "text": "Read unread mail from the last day" },
    { "connection": "slack", "text": "Post a summary to #ops" }
  ],
  "connections": ["gmail", "slack"],
  "model": "claude-sonnet-5-5"
}

Two plans. The workspace pays.

Start on Free with your own browser and every model. Pro adds unlimited runs, scheduled automations and a team. How credits work

Free

$0 forever

200 credits a month

  • 20 agent runs a month
  • Every model on the picker
  • One private browser
  • Descriptions read into blueprints
Join the waitlist

Pro

$29 a month

2,500 credits a month

  • Unlimited runs
  • Every model on the picker
  • Several browsers, scheduled automations
  • Workspaces with members and roles
  • Priority support
Join the waitlist

What changed lately.

Every change that reaches people is written down the day it ships, in their words. Follow it on the page, by RSS, or as JSON if a program is the one keeping watch.

The whole changelog

From the blog

  1. Oct 23, 2026Introducing Describe
  2. Oct 23, 2026Sign in with Google, Apple, GitHub or email
  3. Oct 23, 2026Workspaces with members, roles and invitations

Get a spot.

Leave an email and we'll write when it's your turn. Free is 20 runs a month with every model; Pro is $29.

Get a spot.

We are letting people in from the list in order. Leave an email and we'll write once, when it's your turn.

Free is 20 runs a month with every model on the picker. No card.