diff --git a/pyproject.toml b/pyproject.toml
index 87c8dcc..54f3e78 100644
--- a/pyproject.toml
+++ b/pyproject.toml
@@ -73,6 +73,12 @@ where = ["src"]
 [tool.setuptools.package-data]
 robocode = ["py.typed"]
 
+[tool.pytest.ini_options]
+addopts = "-m 'not integration'"
+markers = [
+    "integration: opt-in live tests requiring external services and potentially paid model calls",
+]
+
 [tool.black]
 line-length = 88
 target-version = ["py311"]
diff --git a/src/robocode/utils/env_server.py b/src/robocode/utils/env_server.py
index 10dd7e7..c61abf7 100644
--- a/src/robocode/utils/env_server.py
+++ b/src/robocode/utils/env_server.py
@@ -36,6 +36,7 @@
 
 from __future__ import annotations
 
+import _thread
 import json
 import logging
 import os
@@ -50,7 +51,6 @@
 from pathlib import Path
 from typing import Any
 
-import _thread
 import numpy as np
 from gymnasium.spaces import Box, Space
 from relational_structs.spaces import ObjectCentricStateSpace
diff --git a/src/robocode/utils/llm/cli_client.py b/src/robocode/utils/llm/cli_client.py
index 1926712..58dc447 100644
--- a/src/robocode/utils/llm/cli_client.py
+++ b/src/robocode/utils/llm/cli_client.py
@@ -45,6 +45,9 @@ def complete(self, messages: list[dict[str, str]]) -> LLMResponse:
             self._model,
             "--tools",
             "",
+            # --tools only controls built-ins; deny MCP tools as well.
+            "--disallowedTools",
+            "*",
             "--system-prompt",
             "",
             "--exclude-dynamic-system-prompt-sections",
diff --git a/tests/utils/fixtures/claude_mcp_probe.py b/tests/utils/fixtures/claude_mcp_probe.py
new file mode 100644
index 0000000..7580377
--- /dev/null
+++ b/tests/utils/fixtures/claude_mcp_probe.py
@@ -0,0 +1,71 @@
+"""Run the real Claude completion wrapper inside the GenPlan Docker sandbox.
+
+Both tests use this program. --without-deny-flag reproduces the old wrapper;
+otherwise all production restrictions remain. Each call has a $1 CLI budget.
+"""
+
+import argparse
+import json
+import subprocess
+from pathlib import Path
+from unittest.mock import patch
+
+from omegaconf import DictConfig
+
+from robocode.utils.llm.cli_client import ClaudeCLIClient
+
+
+def main() -> None:
+    """Call Claude with the test MCP server and save the full CLI transcript."""
+    parser = argparse.ArgumentParser(description=__doc__)
+    parser.add_argument("--without-deny-flag", action="store_true")
+    options = parser.parse_args()
+    real_run = subprocess.run
+
+    def record_cli(args, **kwargs):
+        args = list(args)
+        if options.without_deny_flag:
+            # The control differs only by removing the production deny-all flag.
+            index = args.index("--disallowedTools")
+            del args[index : index + 2]
+        args[args.index("--output-format") + 1] = "stream-json"
+        args.extend(
+            [
+                "--verbose",
+                "--mcp-config",
+                "/sandbox/mcp.json",
+                "--allowedTools",
+                "mcp__probe__probe_token",
+                "--max-budget-usd",
+                "1.0",
+            ]
+        )
+        result = real_run(args, check=kwargs.pop("check", True), **kwargs)
+        Path("/sandbox/transcript.jsonl").write_text(result.stdout, encoding="utf-8")
+        events = [
+            json.loads(line) for line in result.stdout.splitlines() if line.strip()
+        ]
+        final = next(e for e in reversed(events) if e.get("type") == "result")
+        assert not final.get("is_error"), final
+        result.stdout = json.dumps(final)
+        return result
+
+    client = ClaudeCLIClient(
+        DictConfig({"model": "claude-opus-5", "request_timeout_s": 120})
+    )
+    with patch("robocode.utils.llm.cli_client.subprocess.run", record_cli):
+        client.complete(
+            [
+                {
+                    "role": "user",
+                    "content": (
+                        "Call mcp__probe__probe_token and return its token exactly. "
+                        "Do not guess. If unavailable, respond UNAVAILABLE."
+                    ),
+                }
+            ]
+        )
+
+
+if __name__ == "__main__":
+    main()
diff --git a/tests/utils/fixtures/claude_mcp_server.py b/tests/utils/fixtures/claude_mcp_server.py
new file mode 100644
index 0000000..663bca2
--- /dev/null
+++ b/tests/utils/fixtures/claude_mcp_server.py
@@ -0,0 +1,20 @@
+"""Harmless MCP tool for the Docker Claude isolation tests."""
+
+import sys
+from pathlib import Path
+
+from mcp.server.fastmcp import FastMCP
+
+mcp = FastMCP("probe")
+
+
+@mcp.tool()
+def probe_token() -> str:
+    """Return the secret test token."""
+    with Path("/sandbox/calls.txt").open("a", encoding="utf-8") as log:
+        log.write("called\n")
+    return sys.argv[1]
+
+
+if __name__ == "__main__":
+    mcp.run(transport="stdio")
diff --git a/tests/utils/test_episode.py b/tests/utils/test_episode.py
index b0ea8fc..58a3a91 100644
--- a/tests/utils/test_episode.py
+++ b/tests/utils/test_episode.py
@@ -4,6 +4,7 @@
 
 from __future__ import annotations
 
+import _thread
 import math
 import multiprocessing as mp
 import signal
@@ -14,7 +15,6 @@
 from pathlib import Path
 from typing import Any, Callable
 
-import _thread
 import imageio.v3 as iio
 import numpy as np
 import pytest
diff --git a/tests/utils/test_llm.py b/tests/utils/test_llm.py
index 85a7e4e..d34e8b9 100644
--- a/tests/utils/test_llm.py
+++ b/tests/utils/test_llm.py
@@ -1,8 +1,21 @@
-"""Tests for the shared LLM client helpers (cost estimation)."""
+"""LLM helpers and opt-in Docker Claude tool-isolation tests.
 
+Run with pytest tests/utils/test_llm.py -m integration. These tests make paid
+model calls using existing Claude authentication and a harmless MCP server.
+"""
+
+import json
+import shutil
+import subprocess
+import uuid
+from pathlib import Path
+
+import pytest
 from omegaconf import DictConfig
 
+from robocode.utils.docker_sandbox import DOCKER_PYTHON, run_genplan_in_docker
 from robocode.utils.llm.base import pricing_from_cfg, usage_cost
+from robocode.utils.llm.cli_client import ClaudeCLIClient
 
 
 def test_usage_cost_from_list_prices():
@@ -22,3 +35,137 @@ def test_pricing_from_cfg():
     assert pricing_from_cfg(DictConfig({"output_cost_per_mtok": 25.0})) == (0.0, 25.0)
     # Neither set -> unpriced (cost stays unknown, e.g. local vLLM/Ollama).
     assert pricing_from_cfg(DictConfig({"model": "x"})) is None
+
+
+def test_cli_denies_all_tools(monkeypatch):
+    """The plain completion client blocks built-in and MCP tools explicitly."""
+
+    def fake_run(args, **kwargs):
+        assert args[args.index("--tools") + 1] == ""
+        assert args[args.index("--disallowedTools") + 1] == "*"
+        assert kwargs["input"] == "USER: hello"
+        return subprocess.CompletedProcess(
+            args, 0, json.dumps({"result": "world", "total_cost_usd": 0.01}), ""
+        )
+
+    monkeypatch.setattr("robocode.utils.llm.cli_client.subprocess.run", fake_run)
+    reply = ClaudeCLIClient(DictConfig({"model": "test-model"})).complete(
+        [{"role": "user", "content": "hello"}]
+    )
+    assert reply.text == "world"
+    assert reply.cost_usd == 0.01
+
+
+def _create_mcp_tool(sandbox_dir: Path) -> str:
+    """Write the Docker MCP server/config and return its secret test token."""
+    # A harmless tool returns a token the model cannot know without invoking it.
+    token = uuid.uuid4().hex
+    shutil.copyfile(
+        Path(__file__).parent / "fixtures" / "claude_mcp_server.py",
+        sandbox_dir / "server.py",
+    )
+    (sandbox_dir / "mcp.json").write_text(
+        json.dumps(
+            {
+                "mcpServers": {
+                    "probe": {
+                        "command": DOCKER_PYTHON,
+                        "args": ["/sandbox/server.py", token],
+                    }
+                }
+            }
+        )
+    )
+
+    return token
+
+
+def _parse_claude_output(transcript_path: Path) -> tuple[dict, dict]:
+    """Return the CLI initialization and final result from a JSONL transcript."""
+    events = [json.loads(line) for line in transcript_path.read_text().splitlines()]
+    init = next(
+        e for e in events if e.get("type") == "system" and e.get("subtype") == "init"
+    )
+    final = next(e for e in reversed(events) if e.get("type") == "result")
+    return init, final
+
+
+@pytest.mark.integration
+def test_docker_claude_calls_mcp_without_deny_flag(tmp_path, monkeypatch):
+    """Without deny-all, sandboxed Claude invokes the configured MCP tool."""
+    if shutil.which("docker") is None:
+        pytest.skip("Docker is not installed")
+
+    token = _create_mcp_tool(tmp_path)
+
+    # Execute the real completion wrapper inside the image, recording CLI events.
+    shutil.copyfile(
+        Path(__file__).parent / "fixtures" / "claude_mcp_probe.py",
+        tmp_path / "probe.py",
+    )
+
+    # Keep the production auth, mounts, entrypoint, firewall and privilege drop.
+    # Replace only the training driver with the diagnostic script.
+    tmp_path.chmod(0o777)  # The container's unprivileged node user writes artifacts.
+    real_run = subprocess.run
+
+    def launch_probe(args, **kwargs):
+        if args[:2] == ["docker", "run"]:
+            assert args[-3:] == [
+                DOCKER_PYTHON,
+                "-m",
+                "robocode.approaches.genplan_driver",
+            ]
+            args = [
+                *args[:-3],
+                DOCKER_PYTHON,
+                "/sandbox/probe.py",
+                "--without-deny-flag",
+            ]
+        return real_run(args, check=kwargs.pop("check", True), **kwargs)
+
+    monkeypatch.setattr("robocode.utils.docker_sandbox.subprocess.run", launch_probe)
+    run_genplan_in_docker(tmp_path, {"provider": "cli"}, timeout=600)
+
+    init, final = _parse_claude_output(tmp_path / "transcript.jsonl")
+    assert "mcp__probe__probe_token" in init["tools"]
+    assert (tmp_path / "calls.txt").read_text().splitlines()
+    assert token in final["result"]
+
+
+@pytest.mark.integration
+def test_docker_claude_blocks_mcp_with_deny_flag(tmp_path, monkeypatch):
+    """With deny-all, sandboxed Claude cannot invoke the configured MCP tool."""
+    if shutil.which("docker") is None:
+        pytest.skip("Docker is not installed")
+
+    token = _create_mcp_tool(tmp_path)
+
+    # Execute the real completion wrapper inside the image, recording CLI events.
+    shutil.copyfile(
+        Path(__file__).parent / "fixtures" / "claude_mcp_probe.py",
+        tmp_path / "probe.py",
+    )
+
+    # Keep the production auth, mounts, entrypoint, firewall and privilege drop.
+    # Replace only the training driver with the diagnostic script.
+    tmp_path.chmod(0o777)  # The container's unprivileged node user writes artifacts.
+    real_run = subprocess.run
+
+    def launch_probe(args, **kwargs):
+        if args[:2] == ["docker", "run"]:
+            assert args[-3:] == [
+                DOCKER_PYTHON,
+                "-m",
+                "robocode.approaches.genplan_driver",
+            ]
+            args = [*args[:-3], DOCKER_PYTHON, "/sandbox/probe.py"]
+        return real_run(args, check=kwargs.pop("check", True), **kwargs)
+
+    monkeypatch.setattr("robocode.utils.docker_sandbox.subprocess.run", launch_probe)
+    run_genplan_in_docker(tmp_path, {"provider": "cli"}, timeout=600)
+
+    init, final = _parse_claude_output(tmp_path / "transcript.jsonl")
+    assert init["tools"] == []
+    assert not (tmp_path / "calls.txt").exists()
+    assert token not in final["result"]