As more teams hand browser work to AI agents, the browser itself becomes the weak point. A regular browser was never built to keep separate accounts, sessions, and browser identities apart, so accounts end up sharing the same cookies and fingerprint, and one session can overwrite another.
The fix is isolated browser profiles, and MCP makes connecting agents to them simple. Unlike traditional APIs that require custom integration for each service's endpoints, MCP (Model Context Protocol) gives AI applications one standard way to connect to external tools.
Incogniton's MCP server connects AI agents to your isolated browser profiles. An agent can create and organize profiles, proxies, groups, and tags, and launch profiles with their saved sessions, each keeping its own cookies, fingerprint, and proxy.
What is the Incogniton MCP server?
The Incogniton MCP server connects your AI client to Incogniton's isolated browser profiles through the Model Context Protocol (MCP), an open standard that lets AI models call external tools. It works with Claude Desktop, Claude Code, Cursor, Windsurf, VS Code, and Codex CLI.
You describe a task in plain language, and your client can carry it out using Incogniton's profile-management tools. You can use it to manage profiles, assign proxies and groups, and launch profiles with their saved sessions and configured startup URLs.
That covers profile setup as well as launch management. Ask the agent to create three profiles for a new client, move them into a group named after that client, and assign each one a German proxy from your proxy library. The agent handles the entire setup as one workflow, saving you from configuring each profile, group, and proxy separately.
What an AI agent can do in Incogniton
Pairing an AI agent with Incogniton's isolated profiles lets you manage separate accounts without mixing their sessions. You pick the profiles and describe the task, and the agent can manage and launch them from any supported client, such as Claude Desktop, Cursor, or Codex CLI.
This becomes more valuable as your profile count grows. At 50 or 200 profiles, creating, organizing, and updating each one by hand eats into the week. Instead, you can give the agent one instruction and let it handle supported management operations across multiple profiles. Profile setup is one of the first places where this saves time.
Build profile batches from one request
Creating a new profile involves several steps, such as giving it a name, adding it to a group, applying tags, and assigning a proxy. Repeating those steps across a batch quickly becomes tedious. The agent can create or edit up to 30 profiles in one call and clone a profile up to 20 times. Larger batches need several requests, since anything past the limit is dropped without an error.
In Claude Desktop, for example, you could ask it to "create 15 profiles named Store-01 to Store-15, add them to the Client B group, and tag them EU." If one profile already has the configuration you want, the agent can clone it instead.
Review your proxy library
Once your profiles are set up, you can use the agent to review your proxy library. It can list each proxy's name, host, type, location, and whether it rotates, which helps you confirm that every market you work in has the proxies it needs.
For example, in a coding client like Cursor, you could ask it to "list my proxies by location and show which ones rotate." You get a quick overview instead of scrolling through the library manually. To verify whether a proxy connects and check its exit IP, use the proxy-testing feature in Incogniton or ask Cas to check the proxy from inside Incogniton.
Launch profiles with their saved sessions
Instead of launching each profile by hand, you can ask the agent to launch or stop profiles for you.
From a terminal client like Codex CLI, for example, you could ask it to "launch the profiles in Client A." The desktop app opens each profile with its saved session, so you can pick up where you left off.
If you want a profile to open on a specific page, configure its startup URL in the app or ask Cas. Subsequent launches will open that page automatically. Note that the agent can stop only profiles launched through MCP, not profiles started manually or through the REST API.
A launch can also hand the browser to another tool. Once a profile is running, the agent can check its launch status, which returns a local CDP (Chrome DevTools Protocol) address. Puppeteer, Playwright, or a browser MCP server can attach to that address and control the profile with its saved session already loaded. This works when your AI client runs on the same computer as Incogniton and Allow browser remote control is turned on.
Make bulk changes carefully
Bulk edits through MCP take effect immediately and can't be undone through the MCP server, so check your instructions and review the results before applying changes across a large batch. The agent can't delete profiles or restore them from trash or backups.
If you delete a profile in the Incogniton app, it remains in the trash for seven days, where you can restore it. For profile deletion or recovery, use the Incogniton app or Cas.
Common use cases for AI agents in isolated profiles
Incogniton's MCP server helps with workflows involving separate accounts, sessions, and locations. It sets up and launches the isolated profiles, while the browsing itself can be handled by you, Cas inside Incogniton, or a separate browser-automation script. Here are five workflows where that separation can be useful.
Lead generation and prospect research
Create separate profiles for research streams organized by industry, region, or client. Assign the appropriate proxies and launch the profiles with their saved sessions. You can then conduct research yourself or use a separate browser-automation workflow to collect information from directories, company websites, and public social pages.
E-commerce and price monitoring
Set up profiles with proxies for the markets you want to monitor. Launching them with the appropriate sessions and startup URLs prepares your environments for price, stock, and product-listing checks. A separate browsing tool or script is needed to collect that data.
Social media management
Agencies often manage multiple client accounts. Organizing profiles by client and keeping their sessions separate helps teams manage those environments. Actual posting, monitoring, and reporting require the platform's own tools, Cas where supported, or a separate automation workflow. Always check each platform's rules on account management and automation.
QA and testing
Create separate profiles to test how a site behaves across different browser configurations, sessions, and proxy locations. MCP can prepare and launch those environments, but running user journeys and collecting test results requires a separate browser-automation method.
Affiliate and traffic workflows
Organize profiles by campaign or destination to keep session data and browser configurations separate. MCP can help create, configure, and launch those environments, while campaign checks and other repetitive browser tasks require you or a separate automation tool.
Browser automation still needs oversight. Whether you're browsing manually, using Cas, or running a separate automation script, tools can misread page content or take an unexpected action. Review the results, especially when setting up a new workflow.
How to connect AI agents to Incogniton using MCP
Incogniton V5 added MCP support, and the app generates a ready-to-paste configuration for each supported client, so there's no code to write.
Before you start
- Incogniton V5 installed, open, and logged in to the account whose profiles the agent will manage
- Node.js 18 or newer, since the configuration starts the MCP server with npx
- One of the 6 supported AI clients: Claude Desktop, Claude Code, Cursor, Windsurf, VS Code, or Codex CLI
Connect your client
- Generate a token. In Incogniton, go to Settings > Automation, scroll to MCP token, and click Generate MCP Token. The token is shown only once, so copy it before leaving the page. Treat it like a password, since it gives your AI client access to your profiles. You can revoke it at any time from the Active tokens list.
- Copy the configuration. Under Install in your AI client, select your client from the AI client dropdown, then click Copy configuration. Generate the token first so it's filled into the block rather than the YOUR_MCP_TOKEN placeholder.
- Paste it into your client. Follow the location shown under the dropdown. In Claude Desktop, that's Settings > Developer > Edit Config.
- Restart your client. This loads the new connection.
- Verify the connection. Ask your agent to list your Incogniton profiles. If it returns them, the connection works. Then try launching one, and the desktop app will open it within a few seconds.
Once it works, try any of the example requests from earlier in this guide.
Allow browser remote control. Leave this on (the default) in the same Automation tab. Managing and launching profiles works without it, but Puppeteer, Playwright, or a browser MCP server needs it to attach to the profiles the agent launches. Cas also relies on it for its browser actions.
Configuration reference
To connect Incogniton to an MCP-compatible client, you add its server configuration to the client's MCP configuration file. For example, in VS Code, you can add the Incogniton server to .vscode/mcp.json.
{
"servers": {
"incogniton": {
"type": "stdio",
"command": "npx",
"args": ["-y", "@incogniton/mcp"],
"env": {
"INCOGNITON_MCP_TOKEN": "YOUR_MCP_TOKEN"
}
}
}
}
Now let's go through each line of the JSON snippet :
The stdio type tells VS Code to communicate with the server through standard input and output, and the env block passes your MCP token so the server can authenticate the connection.
Other clients use a different file location and format, so copy the block Incogniton generates for your client rather than adapting this one.
Best practices for running AI agents in isolated browser profiles
An agent can change dozens of profiles in one request, and MCP edits can't be undone. These practices keep each workflow contained and help you catch mistakes while they're still small.
Use one profile per workflow
Keep each workflow in its own browser profile. If 2 clients or unrelated workflows need to run independently, create separate profiles rather than sharing one. This keeps each workflow's browser state isolated and prevents one from inheriting cookies, sessions, or other state from another.
Keep fingerprint settings at their defaults
The agent can change fingerprint settings when it creates or edits profiles, including user agent, screen resolution, platform, hardware values, and canvas, audio, and WebGL noise. A poorly chosen change can make a profile less consistent rather than more realistic, creating the same inconsistencies you were trying to avoid. Tell the agent to leave these settings, and anything marked advanced, at their defaults unless you know what you're changing. If you're unsure about a setting, refer to Incogniton's documentation first.
Separate human and agent-managed profiles
Keep profiles managed by AI agents separate from those used for manual work. Create a dedicated group for agent-managed profiles and label it clearly so you know which ones are controlled through MCP. This is especially useful in shared accounts, where other team members could otherwise open or modify a profile the agent is managing. Keep in mind that an MCP token can access everything the account can see, so treat a token on a shared account like a shared key.
Start with small MCP tasks
Start by giving an MCP-connected agent a small, low-risk task and supervise its first few runs. This lets you see how it handles your profiles and catch unexpected behavior before giving it a larger workflow. Once it consistently performs the task as expected, gradually increase the scope of what you delegate to it.
Verify profiles before MCP tasks
Before giving an MCP agent a recurring task, check that the profiles it will use have the expected proxy, browser version, and other settings. This gives the agent a known configuration to work with and helps avoid problems caused by changes made between runs.
Conclusion
AI agents can take over a lot of the setup work behind browser-based workflows, but separate accounts and sessions are hard to manage in a shared browser environment. Incogniton's MCP server lets compatible AI clients create, organize, and launch isolated browser profiles, so each account keeps its own cookies, fingerprint, proxy, and logged-in session.
MCP also removes much of the plumbing and glue code normally needed to connect an AI client to an external service. Connect your client, describe what you need in plain language, and let the agent use the tools exposed by Incogniton's MCP server.
The key is the combination. MCP provides the connection, while isolated profiles keep each account's browser environment separate. Start with small workflows, then expand what you delegate as you become comfortable with the setup.
