Merge browser runtime deployment changes

.env.example

@@ -13,10 +13,17 @@ TELEGRAM_ALLOWED_USERS=
 TELEGRAM_HOME_CHANNEL=
 
 BROWSER_URL=http://browser:9222
-BROWSER_VIEW_URL=
+BROWSER_VIEW_URL=http://localhost:6080
+BROWSER_VIEW_BASE_URL=http://localhost:6081
 
 BROWSER_API_HOST=0.0.0.0
 BROWSER_API_PORT=8088
 BROWSER_USE_RPC_URL=http://browser:8787/run
 BROWSER_USE_RPC_TIMEOUT=900
-BROWSER_API_MAX_CONCURRENCY=2
+BROWSER_API_MAX_CONCURRENCY=2
+BROWSER_USE_ISOLATION_MODE=docker-per-principal
+BROWSER_RUNTIME_IMAGE=browser-use-browser-runtime:latest
+BROWSER_RUNTIME_NETWORK=browser-net
+BROWSER_RUNTIME_TTL_SECONDS=900
+BROWSER_RUNTIME_START_TIMEOUT=45
+BROWSER_RUNTIME_ENABLE_UI=true

.gitea/workflows/deploy.yml

@@ -0,0 +1,22 @@
+name: Deploy to BrowserUse VPS
+
+on:
+  push:
+    branches:
+      - feature/api-for-subagent
+  workflow_dispatch:
+
+jobs:
+  deploy:
+    runs-on: deploy-vps
+    env:
+      DEPLOY_DIR: /home/BrowserUse-vps/apps/BrowserUse_and_ComputerUse_skills
+      DEPLOY_BRANCH: feature/api-for-subagent
+      HEALTH_URL: http://127.0.0.1:8088/health
+    steps:
+      - name: Deploy Docker Compose stack
+        shell: bash
+        run: |
+          set -Eeuo pipefail
+          cd "$DEPLOY_DIR"
+          bash scripts/deploy_vps.sh

api/Dockerfile

@@ -5,6 +5,10 @@ ENV PYTHONUNBUFFERED=1
 
 WORKDIR /app
 
+RUN apt-get update \
+	&& apt-get install -y --no-install-recommends docker.io \
+	&& rm -rf /var/lib/apt/lists/*
+
 COPY requirements.txt /app/requirements.txt
 RUN pip install --no-cache-dir uv \
 	&& uv pip install --system --no-cache-dir -r /app/requirements.txt

api/clients/browser_rpc_client.py

@@ -10,12 +10,13 @@ class BrowserRpcClient:
         self._rpc_url = rpc_url
         self._session = session
 
-    async def run(self, task: str, timeout_sec: float) -> dict[str, Any]:
+    async def run(self, task: str, timeout_sec: float, rpc_url: str | None = None) -> dict[str, Any]:
         payload = {"task": task}
         timeout = aiohttp.ClientTimeout(total=timeout_sec)
+        target_url = rpc_url or self._rpc_url
 
         try:
-            async with self._session.post(self._rpc_url, json=payload, timeout=timeout) as response:
+            async with self._session.post(target_url, json=payload, timeout=timeout) as response:
                 if response.status >= 400:
                     body = await response.text()
                     raise BrowserRpcError(f"RPC HTTP: {response.status}: {body}")

api/clients/browser_rpc_contracts.py

@@ -5,4 +5,4 @@ class BrowserRpcError(RuntimeError): ...
 
 
 class BrowserRpcRunner(Protocol):
-    async def run(self, task: str, timeout_sec: float) -> dict[str, Any]: ...
+    async def run(self, task: str, timeout_sec: float, rpc_url: str | None = None) -> dict[str, Any]: ...

api/services/browser_runtime_manager.py

@@ -0,0 +1,464 @@
+"""Provision isolated browser-use Docker runtimes for API runs."""
+
+from __future__ import annotations
+
+import hashlib
+import json
+import logging
+import os
+import re
+import subprocess
+import tempfile
+import threading
+import time
+from dataclasses import dataclass
+from pathlib import Path
+from typing import Any
+from urllib import request
+
+logger = logging.getLogger(__name__)
+
+_DEFAULT_SHARED_CDP_URL = "http://browser:9222"
+_DEFAULT_SHARED_RPC_URL = "http://browser:8787/run"
+_DEFAULT_RUNTIME_IMAGE = "browser-use-browser-runtime:latest"
+_DEFAULT_RUNTIME_NETWORK = "browser-net"
+_DEFAULT_TTL_SECONDS = 900
+_DEFAULT_START_TIMEOUT = 45
+_DEFAULT_ENABLE_UI = True
+_REGISTRY_LOCK = threading.Lock()
+_VIEW_URL_CACHE_LOCK = threading.Lock()
+_VIEW_URL_CACHE: dict[str, Any] = {"value": "", "expires_at": 0.0}
+
+
+@dataclass(frozen=True)
+class BrowserRuntimeConfig:
+    mode: str
+    runtime_image: str
+    runtime_network: str
+    runtime_ttl_seconds: int
+    runtime_start_timeout: int
+    shared_cdp_url: str
+    enable_ui: bool
+
+
+def _state_dir() -> Path:
+    return Path(os.getenv("BROWSER_RUNTIME_STATE_DIR", "/tmp/browser-use-api"))
+
+
+def _registry_path() -> Path:
+    return _state_dir() / "docker_runtimes.json"
+
+
+def _as_int(value: Any, default: int) -> int:
+    try:
+        return max(1, int(value))
+    except (TypeError, ValueError):
+        return default
+
+
+def _as_bool(value: Any, default: bool) -> bool:
+    if value is None:
+        return default
+    if isinstance(value, bool):
+        return value
+    return str(value).strip().lower() in {"1", "true", "yes", "on"}
+
+
+def get_browser_runtime_config() -> BrowserRuntimeConfig:
+    mode = str(os.getenv("BROWSER_USE_ISOLATION_MODE", "shared")).strip().lower()
+    if mode not in {"shared", "docker-per-principal", "docker-per-task"}:
+        logger.warning("Unknown browser-use isolation mode %r; falling back to shared", mode)
+        mode = "shared"
+
+    return BrowserRuntimeConfig(
+        mode=mode,
+        runtime_image=os.getenv("BROWSER_RUNTIME_IMAGE", _DEFAULT_RUNTIME_IMAGE).strip()
+        or _DEFAULT_RUNTIME_IMAGE,
+        runtime_network=os.getenv("BROWSER_RUNTIME_NETWORK", _DEFAULT_RUNTIME_NETWORK).strip()
+        or _DEFAULT_RUNTIME_NETWORK,
+        runtime_ttl_seconds=_as_int(
+            os.getenv("BROWSER_RUNTIME_TTL_SECONDS"),
+            _DEFAULT_TTL_SECONDS,
+        ),
+        runtime_start_timeout=_as_int(
+            os.getenv("BROWSER_RUNTIME_START_TIMEOUT"),
+            _DEFAULT_START_TIMEOUT,
+        ),
+        shared_cdp_url=os.getenv("BROWSER_URL", _DEFAULT_SHARED_CDP_URL).strip()
+        or _DEFAULT_SHARED_CDP_URL,
+        enable_ui=_as_bool(
+            os.getenv("BROWSER_RUNTIME_ENABLE_UI"),
+            _DEFAULT_ENABLE_UI,
+        ),
+    )
+
+
+def resolve_isolation_owner(
+    mode: str,
+    task_id: str | None,
+    metadata: dict[str, Any] | None = None,
+    thread_id: str | None = None,
+) -> str:
+    if mode == "docker-per-task":
+        return (task_id or "default").strip() or "default"
+
+    metadata = metadata or {}
+    for key in ("user_id", "session_id"):
+        value = metadata.get(key)
+        if value not in (None, ""):
+            return str(value).strip() or "default"
+
+    return (thread_id or task_id or "default").strip() or "default"
+
+
+def hash_runtime_owner(owner: str) -> str:
+    return hashlib.sha256(owner.encode("utf-8")).hexdigest()[:16]
+
+
+def _normalize_browser_view_base_url(raw_url: str) -> str:
+    url = (raw_url or "").strip()
+    if not url:
+        return ""
+    for marker in ("/vnc.html", "/index.html"):
+        idx = url.find(marker)
+        if idx != -1:
+            url = url[:idx]
+            break
+    return url.rstrip("/")
+
+
+def _discover_browser_view_base_url_from_tunnel() -> str:
+    now = time.time()
+    with _VIEW_URL_CACHE_LOCK:
+        cached_value = str(_VIEW_URL_CACHE.get("value", "") or "")
+        expires_at = float(_VIEW_URL_CACHE.get("expires_at", 0.0) or 0.0)
+        if cached_value and now < expires_at:
+            return cached_value
+
+    try:
+        result = _run_docker(["logs", "--tail", "200", "browser-use-tunnel"], check=False)
+        combined = "\n".join(part for part in [result.stdout or "", result.stderr or ""] if part)
+        matches = re.findall(r"https://[^\s\"'<>]+", combined)
+        base_url = _normalize_browser_view_base_url(matches[-1]) if matches else ""
+    except Exception as exc:
+        logger.debug("Failed to discover browser view URL from tunnel logs: %s", exc)
+        base_url = ""
+
+    with _VIEW_URL_CACHE_LOCK:
+        _VIEW_URL_CACHE["value"] = base_url
+        _VIEW_URL_CACHE["expires_at"] = now + (60 if base_url else 10)
+
+    return base_url
+
+
+def get_browser_view_url(
+    task_id: str | None = None,
+    metadata: dict[str, Any] | None = None,
+    thread_id: str | None = None,
+) -> str:
+    base_url = _normalize_browser_view_base_url(
+        os.getenv("BROWSER_VIEW_BASE_URL", "") or os.getenv("BROWSER_VIEW_URL", "")
+    )
+    if not base_url:
+        base_url = _discover_browser_view_base_url_from_tunnel()
+    if not base_url:
+        return ""
+
+    config = get_browser_runtime_config()
+    if config.mode == "shared":
+        return f"{base_url}/vnc.html?path=websockify"
+
+    owner = resolve_isolation_owner(config.mode, task_id, metadata, thread_id)
+    owner_hash = hash_runtime_owner(owner)
+    return f"{base_url}/view/{owner_hash}/vnc.html?path=view/{owner_hash}/websockify"
+
+
+def _shared_rpc_url() -> str:
+    return os.getenv("BROWSER_USE_RPC_URL", _DEFAULT_SHARED_RPC_URL).strip() or _DEFAULT_SHARED_RPC_URL
+
+
+def _runtime_rpc_url(container_name: str) -> str:
+    return f"http://{container_name}:8787/run"
+
+
+def _container_name(owner_hash: str) -> str:
+    return f"browser-use-browser-{owner_hash}"
+
+
+def _volume_name(owner_hash: str) -> str:
+    return f"browser-use-profile-{owner_hash}"
+
+
+def _load_registry() -> dict[str, Any]:
+    path = _registry_path()
+    if not path.exists():
+        return {"runtimes": {}}
+    try:
+        with open(path, "r", encoding="utf-8") as fh:
+            data = json.load(fh) or {}
+        if isinstance(data, dict) and isinstance(data.get("runtimes"), dict):
+            return data
+    except Exception as exc:
+        logger.warning("Failed to read browser-use runtime registry %s: %s", path, exc)
+    return {"runtimes": {}}
+
+
+def _save_registry(payload: dict[str, Any]) -> None:
+    path = _registry_path()
+    path.parent.mkdir(parents=True, exist_ok=True)
+    fd, tmp_path = tempfile.mkstemp(dir=str(path.parent), prefix=".browser_use_", suffix=".tmp")
+    try:
+        with os.fdopen(fd, "w", encoding="utf-8") as fh:
+            json.dump(payload, fh, indent=2, sort_keys=True)
+            fh.flush()
+            os.fsync(fh.fileno())
+        os.replace(tmp_path, path)
+    except Exception:
+        try:
+            os.unlink(tmp_path)
+        except OSError:
+            pass
+        raise
+
+
+def _run_docker(args: list[str], check: bool = True) -> subprocess.CompletedProcess[str]:
+    cmd = ["docker", *args]
+    logger.debug("browser-use docker cmd: %s", " ".join(cmd))
+    result = subprocess.run(
+        cmd,
+        capture_output=True,
+        text=True,
+        timeout=120,
+    )
+    if check and result.returncode != 0:
+        stderr = (result.stderr or result.stdout or "").strip()
+        raise RuntimeError(f"Docker command failed ({' '.join(cmd)}): {stderr}")
+    return result
+
+
+def _ensure_docker_access() -> None:
+    _run_docker(["version"], check=True)
+
+
+def _container_exists(container_name: str) -> bool:
+    result = _run_docker(["inspect", container_name], check=False)
+    return result.returncode == 0
+
+
+def _container_running(container_name: str) -> bool:
+    result = _run_docker(["inspect", "-f", "{{.State.Running}}", container_name], check=False)
+    return result.returncode == 0 and result.stdout.strip().lower() == "true"
+
+
+def _remove_container(container_name: str) -> None:
+    if container_name:
+        _run_docker(["rm", "-f", container_name], check=False)
+
+
+def _volume_exists(volume_name: str) -> bool:
+    result = _run_docker(["volume", "inspect", volume_name], check=False)
+    return result.returncode == 0
+
+
+def _ensure_volume(volume_name: str, owner_hash: str) -> None:
+    if _volume_exists(volume_name):
+        return
+    _run_docker(
+        [
+            "volume",
+            "create",
+            "--label",
+            "browser_use.runtime=true",
+            "--label",
+            f"browser_use.owner_hash={owner_hash}",
+            volume_name,
+        ],
+        check=True,
+    )
+
+
+def _remove_volume(volume_name: str) -> None:
+    if volume_name:
+        _run_docker(["volume", "rm", "-f", volume_name], check=False)
+
+
+def _runtime_env_args(browser_view_url: str, config: BrowserRuntimeConfig) -> list[str]:
+    env: dict[str, str] = {
+        "BROWSER_ENABLE_UI": "true" if config.enable_ui else "false",
+        "BROWSER_DATA_DIR": "/data",
+        "BROWSER_USE_RPC_HOST": "0.0.0.0",
+        "BROWSER_USE_RPC_PORT": "8787",
+    }
+
+    if browser_view_url:
+        env["BROWSER_VIEW_URL"] = browser_view_url
+
+    for key in ("MODEL_DEFAULT", "OPENAI_API_KEY", "OPENAI_BASE_URL"):
+        value = os.getenv(key)
+        if value is not None:
+            env[key] = value
+
+    args: list[str] = []
+    for key, value in env.items():
+        args.extend(["-e", f"{key}={value}"])
+    return args
+
+
+def _start_runtime_container(
+    container_name: str,
+    volume_name: str,
+    owner_hash: str,
+    browser_view_url: str,
+    config: BrowserRuntimeConfig,
+) -> None:
+    _ensure_volume(volume_name, owner_hash)
+    run_args = [
+        "run",
+        "-d",
+        "--name",
+        container_name,
+        "--network",
+        config.runtime_network,
+        "--shm-size",
+        "2g",
+        "--label",
+        "browser_use.runtime=true",
+        "--label",
+        f"browser_use.owner_hash={owner_hash}",
+        "--label",
+        "browser_use.managed_by=browser_runtime_manager",
+        *_runtime_env_args(browser_view_url, config),
+        "-v",
+        f"{volume_name}:/data",
+        config.runtime_image,
+    ]
+    _run_docker(run_args, check=True)
+
+
+def _wait_for_runtime(container_name: str, timeout_seconds: int) -> None:
+    deadline = time.time() + timeout_seconds
+    health_url = f"http://{container_name}:8787/health"
+    last_error = ""
+    while time.time() < deadline:
+        try:
+            with request.urlopen(health_url, timeout=2) as response:
+                if 200 <= response.status < 300:
+                    return
+                last_error = f"HTTP {response.status}"
+        except Exception as exc:
+            last_error = str(exc)
+        time.sleep(1)
+    raise RuntimeError(f"Browser runtime {container_name} did not become ready: {last_error}")
+
+
+def _cleanup_expired_runtimes_locked(registry: dict[str, Any], config: BrowserRuntimeConfig) -> None:
+    now = time.time()
+    runtimes = registry.setdefault("runtimes", {})
+    expired_keys: list[str] = []
+    for runtime_key, entry in list(runtimes.items()):
+        last_used = float(entry.get("last_used", 0) or 0)
+        if not last_used or now - last_used < config.runtime_ttl_seconds:
+            continue
+
+        container_name = str(entry.get("container_name", "") or "")
+        volume_name = str(entry.get("volume_name", "") or "")
+        mode = str(entry.get("mode", "") or "")
+        logger.info("Cleaning expired browser-use runtime %s (%s)", runtime_key, container_name)
+        _remove_container(container_name)
+        if mode == "docker-per-task":
+            _remove_volume(volume_name)
+        expired_keys.append(runtime_key)
+
+    for runtime_key in expired_keys:
+        runtimes.pop(runtime_key, None)
+
+
+def ensure_browser_runtime(
+    task_id: str | None = None,
+    metadata: dict[str, Any] | None = None,
+    thread_id: str | None = None,
+) -> dict[str, str]:
+    config = get_browser_runtime_config()
+    if config.mode == "shared":
+        return {
+            "cdp_url": config.shared_cdp_url,
+            "rpc_url": _shared_rpc_url(),
+            "browser_view": get_browser_view_url(task_id=task_id, metadata=metadata, thread_id=thread_id),
+            "isolation_mode": "shared",
+            "owner_hash": "",
+        }
+
+    _ensure_docker_access()
+    owner = resolve_isolation_owner(config.mode, task_id, metadata, thread_id)
+    owner_hash = hash_runtime_owner(owner)
+    runtime_key = f"{config.mode}:{owner_hash}"
+    container_name = _container_name(owner_hash)
+    volume_name = _volume_name(owner_hash)
+    browser_view_url = get_browser_view_url(task_id=task_id, metadata=metadata, thread_id=thread_id)
+
+    with _REGISTRY_LOCK:
+        registry = _load_registry()
+        _cleanup_expired_runtimes_locked(registry, config)
+
+        if _container_running(container_name):
+            registry.setdefault("runtimes", {})[runtime_key] = {
+                "container_name": container_name,
+                "volume_name": volume_name,
+                "last_used": time.time(),
+                "mode": config.mode,
+                "owner_hash": owner_hash,
+            }
+            _save_registry(registry)
+            return {
+                "cdp_url": f"http://{container_name}:9222",
+                "rpc_url": _runtime_rpc_url(container_name),
+                "browser_view": browser_view_url,
+                "isolation_mode": config.mode,
+                "owner_hash": owner_hash,
+            }
+
+        if _container_exists(container_name):
+            _remove_container(container_name)
+
+        _start_runtime_container(container_name, volume_name, owner_hash, browser_view_url, config)
+        _wait_for_runtime(container_name, config.runtime_start_timeout)
+
+        registry.setdefault("runtimes", {})[runtime_key] = {
+            "container_name": container_name,
+            "volume_name": volume_name,
+            "last_used": time.time(),
+            "mode": config.mode,
+            "owner_hash": owner_hash,
+        }
+        _save_registry(registry)
+
+    return {
+        "cdp_url": f"http://{container_name}:9222",
+        "rpc_url": _runtime_rpc_url(container_name),
+        "browser_view": browser_view_url,
+        "isolation_mode": config.mode,
+        "owner_hash": owner_hash,
+    }
+
+
+def cleanup_browser_runtime(
+    task_id: str | None = None,
+    metadata: dict[str, Any] | None = None,
+    thread_id: str | None = None,
+) -> None:
+    config = get_browser_runtime_config()
+    if config.mode != "docker-per-task":
+        return
+
+    owner = resolve_isolation_owner(config.mode, task_id, metadata, thread_id)
+    owner_hash = hash_runtime_owner(owner)
+    runtime_key = f"{config.mode}:{owner_hash}"
+    container_name = _container_name(owner_hash)
+    volume_name = _volume_name(owner_hash)
+
+    with _REGISTRY_LOCK:
+        registry = _load_registry()
+        _remove_container(container_name)
+        _remove_volume(volume_name)
+        registry.setdefault("runtimes", {}).pop(runtime_key, None)
+        _save_registry(registry)

api/services/task_service.py

@@ -5,6 +5,7 @@ from typing import Any
 from api.clients.browser_rpc_contracts import BrowserRpcError, BrowserRpcRunner
 from api.domain.task_status import TaskStatus
 from api.repositories.task_store import TaskRecord, TaskStore
+from api.services.browser_runtime_manager import cleanup_browser_runtime, ensure_browser_runtime
 
 
 class TaskService:
@@ -108,20 +109,28 @@ class TaskService:
         await self._store.publish(task_id, self._event(task_id, "started", {"status": TaskStatus.running.value}))
 
         async with self._semaphore:
+            runtime: dict[str, str] | None = None
             try:
                 if rec.cancel_requested:
                     await self._store.set_cancelled(task_id)
                     await self._store.publish(task_id, self._event(task_id, "cancelled", {"status": TaskStatus.cancelled.value}))
                     return
 
+                runtime = await asyncio.to_thread(
+                    ensure_browser_runtime,
+                    task_id=task_id,
+                    metadata=rec.metadata,
+                    thread_id=rec.thread_id,
+                )
                 rpc_timeout = float(rec.timeout)
                 if self._rpc_timeout_cap is not None:
                     rpc_timeout = min(rpc_timeout, self._rpc_timeout_cap)
 
                 raw = await asyncio.wait_for(
-                    self._rpc_client.run(task=rec.task, timeout_sec=rpc_timeout),
+                    self._rpc_client.run(task=rec.task, timeout_sec=rpc_timeout, rpc_url=runtime.get("rpc_url")),
                     timeout=float(rec.timeout) + 5,
                 )
+                raw = self._with_runtime_metadata(raw, runtime)
                 success = bool(raw.get("success"))
                 await self._store.set_done(
                     task_id=task_id,
@@ -188,6 +197,16 @@ class TaskService:
                         "status": failed.status.value,
                         "error": failed.error,
                     }))
+            finally:
+                try:
+                    await asyncio.to_thread(
+                        cleanup_browser_runtime,
+                        task_id=task_id,
+                        metadata=rec.metadata,
+                        thread_id=rec.thread_id,
+                    )
+                except Exception:
+                    pass
 
     async def _publish_history_events(self, rec: TaskRecord) -> None:
         for index, item in enumerate(rec.history, start=1):
@@ -225,3 +244,17 @@ class TaskService:
                 normalized.append(event)
         return normalized
 
+    @staticmethod
+    def _with_runtime_metadata(raw: dict[str, Any], runtime: dict[str, str] | None) -> dict[str, Any]:
+        if not isinstance(raw, dict) or not runtime:
+            return raw
+
+        enriched = dict(raw)
+        browser_view = runtime.get("browser_view")
+        if browser_view and not enriched.get("browser_view"):
+            enriched["browser_view"] = browser_view
+        enriched["isolation_mode"] = runtime.get("isolation_mode", "shared")
+        owner_hash = runtime.get("owner_hash")
+        if owner_hash:
+            enriched["owner_hash"] = owner_hash
+        return enriched

api/tests/test_browser_runtime_manager.py

@@ -0,0 +1,97 @@
+from unittest.mock import MagicMock, patch
+
+
+def test_resolve_isolation_owner_prefers_user_id():
+    from api.services.browser_runtime_manager import resolve_isolation_owner
+
+    owner = resolve_isolation_owner(
+        "docker-per-principal",
+        task_id="task-1",
+        metadata={"user_id": "user-7", "session_id": "session-9"},
+        thread_id="thread-1",
+    )
+
+    assert owner == "user-7"
+
+
+def test_resolve_isolation_owner_uses_task_for_per_task_mode():
+    from api.services.browser_runtime_manager import resolve_isolation_owner
+
+    owner = resolve_isolation_owner(
+        "docker-per-task",
+        task_id="task-42",
+        metadata={"user_id": "user-7"},
+        thread_id="thread-1",
+    )
+
+    assert owner == "task-42"
+
+
+def test_hash_runtime_owner_is_stable():
+    from api.services.browser_runtime_manager import hash_runtime_owner
+
+    assert hash_runtime_owner("owner-1") == hash_runtime_owner("owner-1")
+    assert hash_runtime_owner("owner-1") != hash_runtime_owner("owner-2")
+
+
+def test_shared_mode_returns_shared_runtime(monkeypatch):
+    from api.services import browser_runtime_manager
+
+    monkeypatch.setenv("BROWSER_USE_ISOLATION_MODE", "shared")
+    monkeypatch.setenv("BROWSER_URL", "http://shared-browser:9333")
+    monkeypatch.setenv("BROWSER_USE_RPC_URL", "http://shared-browser:8787/run")
+    monkeypatch.setenv("BROWSER_VIEW_BASE_URL", "https://viewer.example.com")
+
+    runtime = browser_runtime_manager.ensure_browser_runtime(
+        task_id="task-1",
+        metadata={"user_id": "user-7"},
+        thread_id="thread-1",
+    )
+
+    assert runtime["cdp_url"] == "http://shared-browser:9333"
+    assert runtime["rpc_url"] == "http://shared-browser:8787/run"
+    assert runtime["browser_view"] == "https://viewer.example.com/vnc.html?path=websockify"
+    assert runtime["isolation_mode"] == "shared"
+
+
+def test_isolated_mode_starts_container(monkeypatch):
+    from api.services import browser_runtime_manager
+
+    monkeypatch.setenv("BROWSER_USE_ISOLATION_MODE", "docker-per-principal")
+    monkeypatch.setenv("BROWSER_RUNTIME_IMAGE", "browser-use-browser-runtime:test")
+    monkeypatch.setenv("BROWSER_RUNTIME_NETWORK", "browser-net")
+    monkeypatch.setenv("BROWSER_VIEW_BASE_URL", "https://viewer.example.com")
+
+    saved_registry = {}
+    docker_calls = []
+
+    def fake_run_docker(args, check=True):
+        docker_calls.append(args)
+        if args[:2] == ["inspect", "-f"]:
+            return MagicMock(returncode=1, stdout="", stderr="")
+        if args[:1] == ["inspect"]:
+            return MagicMock(returncode=1, stdout="", stderr="")
+        return MagicMock(returncode=0, stdout="ok", stderr="")
+
+    with (
+        patch.object(browser_runtime_manager, "_load_registry", return_value={"runtimes": {}}),
+        patch.object(browser_runtime_manager, "_save_registry", side_effect=lambda payload: saved_registry.update(payload)),
+        patch.object(browser_runtime_manager, "_run_docker", side_effect=fake_run_docker),
+        patch.object(browser_runtime_manager, "_wait_for_runtime") as mock_wait,
+    ):
+        runtime = browser_runtime_manager.ensure_browser_runtime(
+            task_id="task-1",
+            metadata={"user_id": "user-7"},
+            thread_id="thread-1",
+        )
+
+    assert runtime["isolation_mode"] == "docker-per-principal"
+    assert runtime["cdp_url"].startswith("http://browser-use-browser-")
+    assert runtime["rpc_url"].startswith("http://browser-use-browser-")
+    assert runtime["rpc_url"].endswith(":8787/run")
+    assert "/view/" in runtime["browser_view"]
+    assert saved_registry["runtimes"]
+    run_commands = [call for call in docker_calls if call[:2] == ["run", "-d"]]
+    assert run_commands
+    assert "browser-use-browser-runtime:test" in run_commands[0]
+    mock_wait.assert_called_once()

api/tests/test_task_service_browser_runtime.py

@@ -0,0 +1,62 @@
+import asyncio
+from typing import Any
+
+
+class FakeRpcClient:
+    def __init__(self) -> None:
+        self.calls: list[dict[str, Any]] = []
+
+    async def run(self, task: str, timeout_sec: float, rpc_url: str | None = None) -> dict[str, Any]:
+        self.calls.append({"task": task, "timeout_sec": timeout_sec, "rpc_url": rpc_url})
+        return {"success": True, "result": "done"}
+
+
+def test_task_service_routes_run_to_browser_runtime(monkeypatch):
+    from api.repositories.task_store import TaskStore
+    from api.services import task_service as task_service_module
+    from api.services.task_service import TaskService
+
+    runtime = {
+        "rpc_url": "http://browser-use-browser-abc:8787/run",
+        "browser_view": "https://viewer.example.com/view/abc/vnc.html?path=view/abc/websockify",
+        "isolation_mode": "docker-per-principal",
+        "owner_hash": "abc",
+    }
+    cleanup_calls = []
+
+    monkeypatch.setattr(task_service_module, "ensure_browser_runtime", lambda **_: runtime)
+    monkeypatch.setattr(task_service_module, "cleanup_browser_runtime", lambda **kwargs: cleanup_calls.append(kwargs))
+
+    async def scenario():
+        rpc_client = FakeRpcClient()
+        service = TaskService(
+            store=TaskStore(),
+            rpc_client=rpc_client,
+            max_concurrency=1,
+            rpc_timeout_cap=30,
+        )
+        rec = await service.create_run(
+            thread_id="thread-1",
+            user_input="open example.com",
+            timeout=60,
+            metadata={"user_id": "user-7"},
+        )
+        done = await service.wait_run(rec.task_id, timeout=2)
+        await service.close()
+        return rpc_client, done
+
+    rpc_client, done = asyncio.run(scenario())
+
+    assert rpc_client.calls == [
+        {
+            "task": "open example.com",
+            "timeout_sec": 30,
+            "rpc_url": "http://browser-use-browser-abc:8787/run",
+        }
+    ]
+    assert done is not None
+    assert done.raw_response is not None
+    assert done.raw_response["browser_view"] == runtime["browser_view"]
+    assert done.raw_response["isolation_mode"] == "docker-per-principal"
+    assert done.raw_response["owner_hash"] == "abc"
+    assert cleanup_calls

browser_env/entrypoint.sh

@@ -7,10 +7,11 @@ XVFB_LOG="/tmp/xvfb.log"
 
 VNC_PORT="${VNC_PORT:-5900}"
 NOVNC_PORT="${NOVNC_PORT:-6080}"
-CHROME_LOCAL_DEBUG_PORT="${CHROME_LOCAL_DEBUG_PORT:-9223}"
+CHROME_LOCAL_DEBUG_PORT="${CHROME_LOCAL_DEBUG_PORT:-${BROWSER_CHROME_DEBUG_PORT:-9223}}"
-CHROME_PUBLIC_DEBUG_PORT="${CHROME_PUBLIC_DEBUG_PORT:-9222}"
+CHROME_PUBLIC_DEBUG_PORT="${CHROME_PUBLIC_DEBUG_PORT:-${BROWSER_CDP_PROXY_PORT:-9222}}"
 BROWSER_USE_RPC_PORT="${BROWSER_USE_RPC_PORT:-8787}"
-CHROME_PROFILE_DIR="${CHROME_PROFILE_DIR:-/src/browser_data}"
+CHROME_PROFILE_DIR="${CHROME_PROFILE_DIR:-${BROWSER_DATA_DIR:-/src/browser_data}}"
+BROWSER_ENABLE_UI="${BROWSER_ENABLE_UI:-true}"
 
 MAX_RESTARTS="${MAX_RESTARTS:-10}"
 RESTART_WINDOW_SEC="${RESTART_WINDOW_SEC:-60}"
@@ -116,19 +117,23 @@ if ! wait_for_x_display 15; then
   exit 1
 fi
 
-start_bg fluxbox
+if [ "$BROWSER_ENABLE_UI" != "false" ]; then
-start_bg x11vnc -display "$DISPLAY" -rfbport "$VNC_PORT" -nopw -listen 0.0.0.0 -xkb -forever -shared
+  start_bg fluxbox
-start_bg websockify --web=/usr/share/novnc/ "$NOVNC_PORT" "localhost:${VNC_PORT}"
+  start_bg x11vnc -display "$DISPLAY" -rfbport "$VNC_PORT" -nopw -listen 0.0.0.0 -xkb -forever -shared
+  start_bg websockify --web=/usr/share/novnc/ "$NOVNC_PORT" "localhost:${VNC_PORT}"
+fi
 start_bg socat "TCP-LISTEN:${CHROME_PUBLIC_DEBUG_PORT},fork,reuseaddr" "TCP:127.0.0.1:${CHROME_LOCAL_DEBUG_PORT}"
 start_bg python3 -u /src/browser_use_runner.py
 
-if ! wait_for_port 127.0.0.1 "$VNC_PORT" 20; then
+if [ "$BROWSER_ENABLE_UI" != "false" ]; then
-  log "fatal: x11vnc did not open port ${VNC_PORT}"
+  if ! wait_for_port 127.0.0.1 "$VNC_PORT" 20; then
-  exit 1
+    log "fatal: x11vnc did not open port ${VNC_PORT}"
-fi
+    exit 1
-if ! wait_for_port 127.0.0.1 "$NOVNC_PORT" 20; then
+  fi
-  log "fatal: websockify did not open port ${NOVNC_PORT}"
+  if ! wait_for_port 127.0.0.1 "$NOVNC_PORT" 20; then
-  exit 1
+    log "fatal: websockify did not open port ${NOVNC_PORT}"
+    exit 1
+  fi
 fi
 if ! wait_for_port 127.0.0.1 "$BROWSER_USE_RPC_PORT" 20; then
   log "fatal: browser-use RPC did not open port ${BROWSER_USE_RPC_PORT}"
@@ -194,4 +199,3 @@ while true; do
   unset CHROME_EXIT
   unset CHROME_PID
 done
-

browser_env/nginx.browser-view.conf

@@ -0,0 +1,46 @@
+events {}
+
+http {
+    resolver 127.0.0.11 ipv6=off;
+
+    map $http_upgrade $connection_upgrade {
+        default upgrade;
+        '' close;
+    }
+
+    server {
+        listen 8080;
+        server_name _;
+
+        location = / {
+            add_header Content-Type text/plain;
+            return 200 "Browser view proxy is running.\n";
+        }
+
+        location / {
+            proxy_http_version 1.1;
+            proxy_set_header Host $host;
+            proxy_set_header Upgrade $http_upgrade;
+            proxy_set_header Connection $connection_upgrade;
+            proxy_set_header X-Forwarded-For $proxy_add_x_forwarded_for;
+            proxy_set_header X-Forwarded-Proto $scheme;
+            proxy_buffering off;
+            proxy_pass http://browser:6080;
+        }
+
+        location ~ ^/view/(?<owner>[a-f0-9]{16})$ {
+            return 302 /view/$owner/vnc.html?path=view/$owner/websockify;
+        }
+
+        location ~ ^/view/(?<owner>[a-f0-9]{16})/(?<rest>.*)$ {
+            proxy_http_version 1.1;
+            proxy_set_header Host $host;
+            proxy_set_header Upgrade $http_upgrade;
+            proxy_set_header Connection $connection_upgrade;
+            proxy_set_header X-Forwarded-For $proxy_add_x_forwarded_for;
+            proxy_set_header X-Forwarded-Proto $scheme;
+            proxy_buffering off;
+            proxy_pass http://browser-use-browser-$owner:6080/$rest$is_args$args;
+        }
+    }
+}

docker-compose.vps.yml

@@ -0,0 +1,14 @@
+services:
+  browser-api:
+    networks:
+      - browser-net
+      - lambdalab_frontend
+
+  browser-view-proxy:
+    networks:
+      - browser-net
+      - lambdalab_frontend
+
+networks:
+  lambdalab_frontend:
+    external: true

docker-compose.yml

@@ -3,6 +3,7 @@ services:
     build:
       context: ./browser_env
       dockerfile: Dockerfile.browser
+    image: browser-use-browser-runtime:latest
     container_name: browser-use-browser
     environment:
       - MODEL_DEFAULT=${MODEL_DEFAULT:-qwen3.5-122b}
@@ -11,8 +12,8 @@ services:
       - BROWSER_USE_RPC_HOST=0.0.0.0
       - BROWSER_USE_RPC_PORT=8787
     ports:
-      - "6080:6080"
+      - "${BROWSER_NOVNC_PUBLISH:-6080:6080}"
-      - "9222:9222"
+      - "${BROWSER_CDP_PUBLISH:-9222:9222}"
     networks:
       browser-net:
         aliases:
@@ -39,11 +40,23 @@ services:
       - BROWSER_API_HOST=0.0.0.0
       - BROWSER_API_PORT=8088
       - BROWSER_API_MAX_CONCURRENCY=2
+      - BROWSER_VIEW_BASE_URL=${BROWSER_VIEW_BASE_URL:-http://localhost:6081}
+      - BROWSER_USE_ISOLATION_MODE=${BROWSER_USE_ISOLATION_MODE:-docker-per-principal}
+      - BROWSER_RUNTIME_IMAGE=${BROWSER_RUNTIME_IMAGE:-browser-use-browser-runtime:latest}
+      - BROWSER_RUNTIME_NETWORK=${BROWSER_RUNTIME_NETWORK:-browser-net}
+      - BROWSER_RUNTIME_TTL_SECONDS=${BROWSER_RUNTIME_TTL_SECONDS:-900}
+      - BROWSER_RUNTIME_START_TIMEOUT=${BROWSER_RUNTIME_START_TIMEOUT:-45}
+      - BROWSER_RUNTIME_ENABLE_UI=${BROWSER_RUNTIME_ENABLE_UI:-true}
+      - MODEL_DEFAULT=${MODEL_DEFAULT:-qwen3.5-122b}
+      - OPENAI_API_KEY=${OPENAI_API_KEY}
+      - OPENAI_BASE_URL=${OPENAI_BASE_URL}
     depends_on:
       browser:
         condition: service_healthy
     ports:
-      - "8088:8088"
+      - "${BROWSER_API_PUBLISH:-8088:8088}"
+    volumes:
+      - /var/run/docker.sock:/var/run/docker.sock
     healthcheck:
       test:
         [
@@ -58,9 +71,34 @@ services:
     networks:
       - browser-net
 
+  browser-view-proxy:
+    image: nginx:alpine
+    container_name: browser-use-view-proxy
+    volumes:
+      - ./browser_env/nginx.browser-view.conf:/etc/nginx/nginx.conf:ro
+    depends_on:
+      browser:
+        condition: service_healthy
+    ports:
+      - "${BROWSER_VIEW_PROXY_PUBLISH:-6081:8080}"
+    restart: always
+    networks:
+      - browser-net
+
+  tunnel:
+    image: cloudflare/cloudflared:latest
+    profiles:
+      - remote
+    container_name: browser-use-tunnel
+    restart: always
+    command: tunnel --protocol http2 --url http://browser-view-proxy:8080 --no-tls-verify
+    networks:
+      - browser-net
+
 volumes:
   browser_profiles:
 
 networks:
   browser-net:
-    driver: bridge
+    name: browser-net
+    driver: bridge

docs/vps-deploy.md

@@ -0,0 +1,133 @@
+# BrowserUse VPS Deployment
+
+This project deploys to `BrowserUse-vps@lambda.coredump.ru` with a Gitea/Forgejo Actions runner installed on the VPS.
+
+The server already has a root-owned `/opt/lambdalab` stack with Caddy on ports `80/443`. Keep this browser service as a separate app under the deploy user home directory, then attach the public-facing containers to the existing `lambdalab_frontend` Docker network through `docker-compose.vps.yml`.
+
+## SSH Access
+
+Add the public SSH key to the VPS user:
+
+```sh
+ssh BrowserUse-vps@lambda.coredump.ru
+mkdir -p ~/.ssh
+chmod 700 ~/.ssh
+printf '%s\n' '<ssh-ed25519 public key>' >> ~/.ssh/authorized_keys
+chmod 600 ~/.ssh/authorized_keys
+```
+
+The fingerprint `SHA256:/XC5ifPX8j+uRyp0Yw2zAl5nteWc3YcHeVHfCG+rhP4` is not enough by itself. `authorized_keys` needs the full public key line that starts with `ssh-ed25519`.
+
+## Initial Server Checkout
+
+Run once on the VPS:
+
+```sh
+mkdir -p ~/apps
+cd ~/apps
+git clone -b feature/api-for-subagent https://git.lambda.coredump.ru/APEX/BrowserUse_and_ComputerUse_skills.git
+cd BrowserUse_and_ComputerUse_skills
+```
+
+Create a server-local `.env` file in the checkout. It is intentionally not committed:
+
+```sh
+OPENAI_API_KEY=...
+OPENAI_BASE_URL=...
+MODEL_DEFAULT=qwen3.5-122b
+BROWSER_VIEW_BASE_URL=https://browser-view.lambda.coredump.ru
+BROWSER_API_PUBLISH=127.0.0.1:8088:8088
+BROWSER_VIEW_PROXY_PUBLISH=127.0.0.1:6081:8080
+BROWSER_NOVNC_PUBLISH=127.0.0.1:6080:6080
+BROWSER_CDP_PUBLISH=127.0.0.1:9222:9222
+```
+
+Then run the first deploy manually:
+
+```sh
+bash scripts/deploy_vps.sh
+curl -fsS http://127.0.0.1:8088/health
+```
+
+The deploy script uses both Compose files by default:
+
+```sh
+docker-compose.yml:docker-compose.vps.yml
+```
+
+`docker-compose.vps.yml` connects `browser-api` and `browser-view-proxy` to the existing external `lambdalab_frontend` network so Caddy can reach them by Docker DNS.
+
+## Domain Binding
+
+The active Caddy config is root-owned at:
+
+```sh
+/opt/lambdalab/caddy/Caddyfile
+```
+
+Add these vhosts to that file from an admin/root account:
+
+```caddyfile
+browser-api.lambda.coredump.ru {
+	reverse_proxy browser-use-api:8088
+}
+
+browser-view.lambda.coredump.ru {
+	reverse_proxy browser-use-view-proxy:8080
+}
+```
+
+Then reload the existing Caddy container from `/opt/lambdalab`:
+
+```sh
+cd /opt/lambdalab
+docker compose exec caddy caddy reload --config /etc/caddy/Caddyfile
+```
+
+DNS must point both subdomains to the VPS public IP `155.212.185.120`. At inspection time, `lambda.coredump.ru` resolved to that IP, while `browser-api.lambda.coredump.ru` and `browser-view.lambda.coredump.ru` did not resolve yet.
+
+## Gitea/Forgejo Runner
+
+Install `act_runner` as the `BrowserUse-vps` user and register it with the repository, organization, or instance runner token:
+
+```sh
+mkdir -p ~/act_runner
+cd ~/act_runner
+./act_runner generate-config > config.yaml
+./act_runner --config config.yaml register \
+  --no-interactive \
+  --instance https://git.lambda.coredump.ru \
+  --token '<runner-registration-token>' \
+  --name BrowserUse-vps \
+  --labels deploy-vps:host
+```
+
+Start it under the same user:
+
+```sh
+cd ~/act_runner
+nohup ./act_runner daemon --config config.yaml > act_runner.log 2>&1 &
+```
+
+Because this account has `sudo: no`, a system-wide service cannot be installed from this user. If an admin enables a user-level systemd service for this account, run the same daemon command from that service instead of `nohup`.
+
+## CI/CD Behavior
+
+The workflow lives at `.gitea/workflows/deploy.yml`.
+
+It runs on:
+
+- push to `feature/api-for-subagent`
+- manual `workflow_dispatch`
+
+The job expects a runner label named `deploy-vps`, registered as `deploy-vps:host`. It enters:
+
+```sh
+/home/BrowserUse-vps/apps/BrowserUse_and_ComputerUse_skills
+```
+
+Then it fetches `origin/feature/api-for-subagent`, resets the tracked checkout to that commit, runs Docker Compose, and verifies:
+
+```sh
+curl -fsS http://127.0.0.1:8088/health
+```

scripts/deploy_vps.sh

@@ -0,0 +1,65 @@
+#!/usr/bin/env bash
+set -Eeuo pipefail
+
+DEPLOY_BRANCH="${DEPLOY_BRANCH:-feature/api-for-subagent}"
+HEALTH_URL="${HEALTH_URL:-http://127.0.0.1:8088/health}"
+COMPOSE_FILES="${COMPOSE_FILES:-docker-compose.yml:docker-compose.vps.yml}"
+
+log() {
+  printf '[deploy] %s\n' "$*"
+}
+
+fail() {
+  printf '[deploy] fatal: %s\n' "$*" >&2
+  exit 1
+}
+
+command -v git >/dev/null 2>&1 || fail "git is not installed"
+command -v docker >/dev/null 2>&1 || fail "docker is not installed"
+command -v curl >/dev/null 2>&1 || fail "curl is not installed"
+docker compose version >/dev/null 2>&1 || fail "docker compose plugin is not available"
+
+[ -d .git ] || fail "current directory is not a git checkout"
+[ -f docker-compose.yml ] || fail "docker-compose.yml not found in current directory"
+[ -f .env ] || fail ".env is missing; create it on the VPS with OPENAI_API_KEY and related runtime settings"
+
+compose_args=()
+IFS=':' read -r -a compose_files <<< "$COMPOSE_FILES"
+for compose_file in "${compose_files[@]}"; do
+  if [ -f "$compose_file" ]; then
+    compose_args+=("-f" "$compose_file")
+  else
+    fail "compose file not found: ${compose_file}"
+  fi
+done
+
+log "fetching origin/${DEPLOY_BRANCH}"
+git fetch --prune origin "+refs/heads/${DEPLOY_BRANCH}:refs/remotes/origin/${DEPLOY_BRANCH}"
+
+log "checking out ${DEPLOY_BRANCH}"
+git checkout -B "$DEPLOY_BRANCH" "origin/$DEPLOY_BRANCH"
+git reset --hard "origin/$DEPLOY_BRANCH"
+
+log "building Docker Compose services"
+docker compose "${compose_args[@]}" build
+
+log "starting Docker Compose stack"
+docker compose "${compose_args[@]}" up -d --remove-orphans
+
+log "current service state"
+docker compose "${compose_args[@]}" ps
+
+log "waiting for API health at ${HEALTH_URL}"
+for attempt in {1..30}; do
+  if curl -fsS "$HEALTH_URL" >/dev/null; then
+    log "API is healthy"
+    exit 0
+  fi
+
+  log "health check failed, retry ${attempt}/30"
+  sleep 2
+done
+
+log "API did not become healthy; browser-api logs follow"
+docker compose "${compose_args[@]}" logs --tail=120 browser-api || true
+fail "health check failed: ${HEALTH_URL}"