dsh-antibrow

A DeepSeek Harness plugin that gives the agent a browser with an identity.

dsh plugin --profile <name> add dsh-antibrow

Why not a plain browser plugin

Every other browser plugin hands the agent a fresh Chromium. That is fine for reading a public page and useless for anything behind a login: the session dies with the process, the fingerprint is a stock automation build, and the traffic leaves from your machine.

The difference shows up the first time an agent has to be someone: check a mailbox, watch a dashboard, keep a marketplace account warm. A browser with no memory has to log in again on every run, and logging in again is exactly what a site treats as suspicious.

Related MCP server: @meshbrow/mcp-server

What it brings

• Unlimited profiles, on the free tier. Not five, not fifty - profiles live on your disk, not on a plan, and creating one costs nothing. Give every account, every marketplace, every persona its own browser and stop reasoning about which cookie jar the agent is holding. Syncing them between machines and the managed residential proxies are the paid parts; the profiles themselves never are.

• 16 tools for the model: sessions, profiles, proxies, page control, live view. All under mcp__antibrow__*, all reachable from Code Mode as ordinary async calls.

• An Android phone, from the same API. One argument - deviceType: 'android' - and the agent is a phone: touch points, mobile client hints, phone viewport, phone GPU. See below.

• Engine-level spoofing. The user agent, platform, screen, fonts, canvas, WebGL and audio are answered by the engine itself. Nothing is injected into the page, so there is no injected script for a detector to find.

• A coherent story, not a pile of overrides. Timezone, locale and the reported network profile are derived from the proxy's real exit, because a US IP that reports Shanghai time is a contradiction a page can check in one line.

• Passkeys survive. WebAuthn credentials are stored with the profile, not in the machine's keychain, so a passkey registered on one run still signs in on the next - and, with sync on, on another computer.

• Watch it work. start_live_view streams the agent's browser to a dashboard, so a long unattended run is something you can look at instead of guess about.

• Identities from real machines. On a paid plan a profile can be minted from a captured real-device fingerprint rather than a generated one - a whole coherent row off one physical machine, not a field-by-field invention.

• Four builds, three operating systems: Windows, macOS, and Linux on both x86_64 and arm64. The same profile runs on any of them.

• Portable, and not only to agents. Export a profile as a file and import it on another machine; leave it unsynced and it is still there, in the desktop app, for a human to open and finish by hand. The agent's browser and yours can be the same browser.

The phone

{ "profile": "shop-mobile", "deviceType": "android", "temporary": true }

Android profiles are built from whole captured devices - the screen, the GPU report and the client hints agree with each other because they came off the same physical phone, and the plugin picks a row rather than assembling one. Three of them ship inside the package, so a free-tier agent can create an Android profile with no network round trip at all.

The device type is fixed when the profile is created and never drifts afterwards: a profile that was a phone stays that phone. Android needs engine 151 or newer - if none is available the launch fails rather than quietly handing you a desktop browser that claims to be a phone.

Measured

On a macOS host, through a US residential proxy, 2026-08-15, engine 151:

The 10% whoer deducts is WebRTC, which had no route to a STUN server on that run. Reproduce all of it with tests/smoke - the harness is in this repository, and it fails loudly rather than printing a number nobody checks.

Install

dsh plugin --profile <name> add dsh-antibrow
export ANTI_DETECT_BROWSER_KEY=<key>
dsh --profile <name>

The tools

launch_browser takes a profile name and creates it on first use. Pass temporary: true for automation work: those profiles are local-only and stay out of the desktop app's list, while still persisting on disk.

Before you rely on it

• One browser at a time on a free key. The concurrency limit is carried by the license and counted per machine, across every application using the same engine. Fan an agent out into parallel sessions only on a plan whose limit covers them, or the extra launches are refused.

• The engine downloads on first launch (190-320 MB depending on platform), into a shared cache directory. The first launch_browser of a fresh install pays for that; later ones do not.

• A persistent identity is a real identity. Two agents driving one profile at once is the same mistake as two people sharing one browser: last one to close wins. Give each its own.

Configuration

The bundle inserts one row, mcp-antibrow, configuring the harness's MCP client against the anti-detect-browser CLI. Override it from your profile's own cordis.patch.yml by targeting that id - a patch replaces the row's whole config, so restate every key you keep:

- id: mcp-antibrow
  config:
    serverName: antibrow
    transport: stdio
    command: npx
    args: ['-y', 'anti-detect-browser@^2.19.1', '--mcp']
    env:
      ANTI_DETECT_BROWSER_KEY: !!js process.env.MY_OWN_VAR
      ANTI_DETECT_BROWSER_CACHE_DIR: /var/lib/antibrow

Turn it off without uninstalling:

- id: mcp-antibrow
  disabled: true

To run the SDK from a checkout instead of npm, use the overlay this package ships:

export ANTIBROW_SDK_CLI=<checkout>/dist/cli.js
dsh --profile <name> --patch node_modules/dsh-antibrow/cordis.patch.local.yml

Compatibility

Verified against dsh 0.1.0-rc.6 and anti-detect-browser 2.19.1, 2026-08-15. DeepSeek Harness is a developer preview and says it will break compatibility; this line is the claim, and it rots without a re-test.

The SDK version floor is real: 2.19.1 is the first release whose close path ends the browser process instead of only dropping the debugging connection. On an older one a closed session leaves the browser running, which keeps that profile's directory locked - the next launch of it dies before the debugging connection is ready - and holds a concurrency slot for the rest of the run.

License

MIT