---
title: "AI Agent for Sales Prospecting: A Private Setup"
description: "Run an AI agent for sales prospecting in private browser profiles: account briefs, lead lists, CRM cleanup and drafts, one profile per rep, approved by the rep."
url: "https://donutbrowser.com/pt/ai-agent-for-sales-prospecting/"
---

1. [Início](https://donutbrowser.com/pt/)

2. [Blog](https://donutbrowser.com/pt/blog)

3. AI Agent for Sales Prospecting: A Private Setup

# AI Agent for Sales Prospecting: A Private Setup

October 14, 2026

ai-agents sales prospecting

**An AI agent for sales prospecting does the browser work around a deal: it researches accounts, builds lead lists, proposes CRM fixes and drafts outreach. Run it in Donut Browser with one isolated profile per rep or sender mailbox, connect your own agent over remote MCP or local CDP, and keep every send and save with a person.**

Donut Browser is the most private anti-detect browser. The desktop app has zero telemetry, needs no account for local use, and keeps profiles, cookies and history on your own disk. It contains no AI model. You bring the agent, such as Claude Code, Cursor or Codex, and Donut gives it browser profiles that do not share a fingerprint, cookies or an IP address.

This guide covers five prospecting workflows: which agent fits each one, how it drives Donut, and where the rep approves. It also covers what LinkedIn’s terms allow, because LinkedIn automation with an AI agent breaks them. For the product view of the same setup, see [AI agents for sales teams](https://donutbrowser.com/use-cases/sales-agents).

![Diagram: Claude Code, Cursor or Codex connect to Donut Browser over remote MCP or local CDP. Each rep and sender mailbox has its own profile with its own fingerprint and proxy. The agent works in the CRM and on company sites, LinkedIn stays with the rep, and ask\_human puts every send or save in front of a person.](https://donutbrowser.com/img/blog/ai-agent-for-sales-prospecting.svg)

## What can an AI agent do for sales prospecting?

An agent fits sales work where the CRM holds the task and the evidence is on other websites. Five workflows cover most of an SDR’s browser time:

| Workflow | Agent that fits | How it drives Donut | Where a person approves |
| --- | --- | --- | --- |
| Account research brief | Claude Desktop or Claude Code | Remote MCP in the rep’s profile | The rep reads the brief before the call |
| Lead list with LinkedIn URLs | Claude Code or Codex | A Playwright script over local CDP | The rep checks the flagged rows |
| CRM cleanup | Claude Code or Cursor | Remote MCP in the rep’s profile | RevOps approves each change |
| Outreach drafts | Claude Desktop | Remote MCP, stops at the draft | The rep edits and sends |
| Scheduled pipeline check | No model: a recipe on a schedule | Donut runs the recorded steps | The rep reads the result |

The [AI agents use case](https://donutbrowser.com/use-cases/ai-agents) covers the same tools for other teams.

## Which AI agent should drive the browser?

Use the agent you already pay for. Donut can write the MCP config for 22 clients, one click per client, among them Claude Desktop, Claude Code, Cursor, VS Code, Codex and Zed. The agent’s model runs with its own provider. Donut adds the browser profiles and the tools to drive them.

There are two ways to connect:

- Remote MCP. The agent calls `https://api.donutbrowser.com/api/mcp` with a `dmk_` credential. The desktop app keeps an outgoing WebSocket to Donut’s relay, so no port opens on your computer. Solo ($29 per month) covers profiles, proxies, groups and the human-in-the-loop tools. Pro ($99 per month) adds launch, navigate, click, type, screenshot and recipes. Setup is in the [Remote MCP docs](https://donutbrowser.com/docs/remote-mcp).

- Local CDP. Turn on the Local API, call `POST /v1/profiles/{id}/run`, and connect Playwright to the CDP port it returns on 127.0.0.1. Coding agents such as Claude Code and Codex write and run these scripts. This path needs Pro. Start with the [quickstart](https://donutbrowser.com/docs/quickstart) and the [Wayfern Playwright wrapper](https://donutbrowser.com/docs/wayfern), which adds human-like mouse movement, typing and scrolling.

Use remote MCP when the agent must decide page by page, as in research, CRM cleanup and drafts. Use local CDP for long, repeated reads such as directories, because a script costs no model tokens per page.

## How does an agent build an account research brief?

A brief collects what a rep needs ten minutes before a call. Give the agent the account name, the CRM record and the meeting time. In the rep’s own profile, it opens the CRM account and the last call notes, then the prospect’s website, pricing page, security page, careers page and recent news. It reads each page, takes screenshots of pages that may change, and returns a short brief with a source link for every claim.

A prompt that works:

```
Use the Donut profile "Rep A". Research Acme before my 2 pm call.
Open the CRM account and the last call notes, acme.com, its pricing,
security and careers pages, and news from the last 90 days.
Return 3 signals with source links, 2 risks and 5 questions.
Do not change the CRM. Do not open linkedin.com.
```

The CRM session is the rep’s own login, so the agent sees only that rep’s records and cannot reach another rep’s accounts.

## How do you build a lead list with LinkedIn URLs?

Build the list from sources that publish the data, then let the rep check LinkedIn by hand. Good sources are trade association directories, conference exhibitor and speaker pages, partner directories and company team pages. Many of them list a LinkedIn URL next to each person or company.

For a directory with hundreds of rows, ask Claude Code or Codex to write a Playwright script with the `wayfern` package. The script runs in a research profile over local CDP, follows each detail link, and writes a CSV with these columns: company, website, name, title, LinkedIn URL from the source page, reason to reach out, proof URL and review status. Rows with a missing field get a flag. The model reads only the rows you ask it to review, so a 500-row directory costs a few prompts.

Pace the script for the site you read. One tab, a pause of several seconds between pages and a daily page cap keep the load close to a person’s. When a CAPTCHA or a login wall appears, the script stops. If the agent works over remote MCP, it calls `request_human_help` instead, and you solve the check in the same profile.

## How does an agent clean up CRM records?

CRM cleanup waits because each field needs a check on another site. The agent can do the checks and leave the decision to RevOps. For each account, it opens the CRM record, the company website and other public sources, and builds a review table with the current value, the proposed value, the source URL and a confidence note. It does not save.

When the table is ready, the agent calls `ask_human` with a summary. The question appears on the Agent page of the Donut desktop app. You approve, reject or reply with notes, and the agent applies only the approved rows. At any point you can pause the agent or take over the profile yourself.

## How should an agent draft outreach?

Let the agent write drafts and stop. For each prospect on the reviewed list, it opens the rep’s email client or sequencing tool in the rep’s profile, writes a draft that cites the trigger and the proof URL, and saves it as a draft. The rep edits and sends.

Put sender mailboxes in their own profiles. Many SDR teams send from secondary domains or from several mailboxes. Email providers watch for sign-ins from new devices and locations, so a mailbox that always opens in the same profile, with the same fingerprint and proxy, looks like one person on one computer. Two mailboxes in one browser share cookies and a device.

## How do scheduled pipeline checks work?

Some checks repeat every week and need no reasoning: open the pipeline view, filter stage-two deals with no activity in 14 days, copy the rows and take a screenshot. Record that as a recipe. You do the task once in the profile’s browser, and Donut writes down the steps. Passwords are never recorded, and a replay uses no model tokens.

A schedule runs the recipe on set weekdays at a set time, once at a date, or at an interval from 15 minutes to 7 days, and a failed run can send you an email or a signed webhook. Schedules start with Solo and run on Donut’s remote fleet, which needs the profile synced without end-to-end encryption. A schedule on your own computer needs Pro. On Pro, your agent can also start the recipe with `run_recipe`, read the extracted rows with `get_recipe_run` and draft next steps for the rep to review.

Recipes and schedules are stored in Donut’s cloud, and each run stores its transcript, extracted values and any screenshots there too. If the CRM rows must stay on your machine, have Claude Code write a Playwright script for the same check and run it with your own scheduler.

## Why does each rep or sender account need its own profile?

Websites link accounts by browser fingerprint, cookies and IP address. When five reps’ sessions share one browser, the agent can act in the wrong account, and every site sees one device behind five people. The [browser fingerprint](https://donutbrowser.com/browser-fingerprinting) alone is enough to connect them.

In Donut, each profile has its own fingerprint and cookie store, its own proxy or WireGuard VPN, WebRTC leak protection, and DNS through the profile’s route. Timezone and language follow the proxy’s exit IP, and Donut blocks the launch when they do not match. See [proxies for anti-detect browsers](https://donutbrowser.com/proxies-for-anti-detect-browsers) to pick an IP for each rep.

One profile per rep also follows LinkedIn’s rules. Its [User Agreement](https://www.linkedin.com/legal/user-agreement) forbids using “another’s account (such as sharing log-in credentials or copying cookies)”. Keep each LinkedIn account in its owner’s profile, and let only that person use it. On the Team plan, profile locks stop two people from opening the same profile at once. Roles limit what each member can change, and group access sets which profiles each member can open.

## What do LinkedIn’s terms say about AI agents?

LinkedIn does not allow automated activity on its website, whichever browser runs it. Its help page on [automated activity](https://www.linkedin.com/help/linkedin/answer/a1340567) says: “we don’t allow the use of third-party software or browser extensions that scrape, modify the appearance of, or automate activity on LinkedIn’s website.” Section 8.2 of the [User Agreement](https://www.linkedin.com/legal/user-agreement) forbids using “bots or other unauthorized automated methods to access the Services, add or download contacts, send or redirect messages”. LinkedIn restricts accounts it catches, and the help page tells their owners to “disable the software or extension that automates activities”.

An AI agent that searches LinkedIn, reads profiles, clicks Connect or sends messages breaks these terms. Donut’s fingerprints, proxies and humanized input change what LinkedIn can measure. They do not change what the terms allow. Behavior also gives agents away: in the 2026 [FP-Agent study](https://arxiv.org/abs/2605.01247), typing, scrolling and mouse behavior separated seven AI browsing agents from human users. [Why websites flag AI agents](https://donutbrowser.com/ai-agent-bot-detection) covers each signal.

A setup that stays inside the terms:

- Keep the agent off linkedin.com, and write that rule into its instructions.

- Take LinkedIn URLs from the directory or company page where you found the lead.

- Let the rep open each LinkedIn profile, decide, and send connection requests and messages by hand.

- Pace the agent on every other site: one tab per profile, several seconds between pages and a daily cap.

## Where does your prospect data go?

Prospect lists and CRM records hold personal data, and each tool in the chain can copy it. With Donut:

- The desktop app sends no telemetry, analytics or crash reports. See [why Donut collects no telemetry](https://donutbrowser.com/no-telemetry).

- Profiles, cookies and history stay on your disk. Cloud sync is off by default, and when you turn it on, end-to-end encryption is available as an option.

- Donut sends nothing to model providers. Your agent sends what it reads to its own provider, under that provider’s terms.

- Remote MCP is not end-to-end encrypted: page text and screenshots pass through Donut’s relay on their way to your agent. The [Remote MCP docs](https://donutbrowser.com/docs/remote-mcp) list what the relay sees and keeps. Over MCP, Donut refuses tools that read local paths or export secrets, and it redacts proxy passwords.

- Recipe runs store their transcripts, extracted values and any screenshots in Donut’s cloud. A local script keeps them on your disk.

- The Agent page keeps the agent console in memory only.

Aside, an AI browser that targets the same sales tasks, says on its [home page](https://aside.com/) that “Nothing is shared with LLM providers.” Section 4 of its [privacy policy](https://aside.com/policy/privacy) says “Hosted AI model providers receive the model-visible context needed to run the task.” Aside can also connect local models through Ollama or LM Studio. Its [privacy help page](https://docs.aside.com/help/privacy) says “Analytics sharing defaults to on.” (All retrieved October 14, 2026.) The full comparison is on the [Aside alternative](https://donutbrowser.com/alternatives/aside) page, and [AI browser privacy](https://donutbrowser.com/ai-browser-privacy) covers what each setup sends to model providers.

## How do you set up an AI SDR browser with Donut?

1. Install Donut on macOS, Windows or Linux. Create one profile per rep and one per sender mailbox, each with a proxy in that person’s region.

2. Sign in on Solo or Pro, open Integrations → Remote MCP in the sidebar, and turn on remote control.

3. Add your agent with one click, or paste the endpoint and a credential into its config:

```
{
  "mcpServers": {
    "donut-browser-remote": {
      "url": "https://api.donutbrowser.com/api/mcp",
      "headers": { "Authorization": "Bearer dmk_…" }
    }
  }
}
```

1. Give the agent its rules: the profile name for each rep, no linkedin.com, stop before any send or save, and `ask_human` for approval.

2. Record a recipe for the weekly pipeline check and put it on a schedule.

Per-client setup is in the [AI agents docs](https://donutbrowser.com/docs/ai-agents). See [pricing](https://donutbrowser.com/pricing) for what each plan includes. The same pattern works for [recruiting](https://donutbrowser.com/ai-agent-for-recruiting) and for [vendor portals and invoices](https://donutbrowser.com/ai-agent-for-operations).

## Frequently asked questions

### What is the best AI agent for sales prospecting?

**Use the agent your team already pays for.** Claude Code, Claude Desktop, Cursor, Codex and other MCP clients all drive Donut Browser. Donut contains no AI model; it gives the agent separate browser profiles and the tools to use them. Coding agents suit list building with scripts, and chat agents suit research briefs and drafts.

### Can an AI agent automate LinkedIn prospecting?

**Not within LinkedIn’s terms.** LinkedIn does not allow third-party software that scrapes or automates activity on its website, and its [User Agreement](https://www.linkedin.com/legal/user-agreement) forbids bots that add contacts or send messages. Let the agent build lists from other sources, and let the rep do every LinkedIn action by hand.

### Why does each sales rep need a separate browser profile?

**Sites link accounts that share a browser fingerprint, cookies or an IP address.** A separate profile keeps each rep’s sessions, device and proxy apart, so the agent works only in that rep’s accounts. LinkedIn’s User Agreement also forbids sharing log-in credentials or copying cookies between people.

### Does Donut Browser send my prospect data to an AI provider?

**No. Donut sends nothing to model providers.** Your agent sends what it reads to its own provider, under that provider’s terms. The desktop app has no telemetry. Remote MCP calls pass through Donut’s relay, and the local API and CDP path stay on your machine.

### Which Donut Browser plan does an AI SDR setup need?

**Pro, for most teams.** Solo ($29 per month) gives remote MCP for profiles, proxies and approvals, without browser control. Pro ($99 per month) adds browser control, the local launch API and recipes that your agent runs on demand. Team ($199 per month for two seats) adds profile locks and roles. See [pricing](https://donutbrowser.com/pricing).

### Can the agent send emails by itself?

**It can, so tell it not to.** Have it save drafts and call `ask_human` with a summary. You approve on the Agent page of the desktop app, and the rep sends. You can pause the agent or take over its profile at any time.

## On this page

- What can an AI agent do for sales prospecting?

- Which AI agent should drive the browser?

- How does an agent build an account research brief?

- How do you build a lead list with LinkedIn URLs?

- How does an agent clean up CRM records?

- How should an agent draft outreach?

- How do scheduled pipeline checks work?

- Why does each rep or sender account need its own profile?

- What do LinkedIn’s terms say about AI agents?

- Where does your prospect data go?

- How do you set up an AI SDR browser with Donut?

- Frequently asked questions

- What is the best AI agent for sales prospecting?

- Can an AI agent automate LinkedIn prospecting?

- Why does each sales rep need a separate browser profile?

- Does Donut Browser send my prospect data to an AI provider?

- Which Donut Browser plan does an AI SDR setup need?

- Can the agent send emails by itself?