diff --git a/src/tether/integrations/__init__.py b/src/tether/integrations/__init__.py
index 7c869c6..b6feecc 100644
--- a/src/tether/integrations/__init__.py
+++ b/src/tether/integrations/__init__.py
@@ -1,8 +1,15 @@
 """Integration framework — connect external tools to tether."""
+from tether.integrations.connector import connect, disconnect
 from tether.integrations.registry import (
     Integration,
     get_integration,
     list_integrations,
 )
 
-__all__ = ["Integration", "get_integration", "list_integrations"]
+__all__ = [
+    "Integration",
+    "connect",
+    "disconnect",
+    "get_integration",
+    "list_integrations",
+]
diff --git a/src/tether/integrations/connector.py b/src/tether/integrations/connector.py
index 8c8b375..8d36d1d 100644
--- a/src/tether/integrations/connector.py
+++ b/src/tether/integrations/connector.py
@@ -1,20 +1,52 @@
 """Connector — install, start, query, and stop integrations."""
 from __future__ import annotations
 
-import json
 import logging
+import os
+import signal
 import subprocess
+import time
+from pathlib import Path
 from typing import Any
 
 import requests
 
-from tether.integrations.registry import Integration, get_integration
+from tether.integrations.registry import get_integration, state_dir
 
 logger = logging.getLogger(__name__)
 
+# Same-process handle cache. NOT relied on for stop — `tether connect up`
+# Popens a child then the CLI process exits, so by the time `tether connect
+# down` runs in a NEW process this dict is empty. The pid file below is what
+# makes stop work across invocations.
 _RUNNING: dict[str, subprocess.Popen] = {}
 
 
+def _pid_file(name: str) -> Path:
+    return state_dir() / f"{name}.pid"
+
+
+def _write_pid(name: str, pid: int) -> None:
+    _pid_file(name).write_text(str(pid))
+
+
+def _read_pid(name: str) -> int | None:
+    try:
+        return int(_pid_file(name).read_text().strip())
+    except (OSError, ValueError):
+        return None
+
+
+def _pid_alive(pid: int) -> bool:
+    try:
+        os.kill(pid, 0)
+    except ProcessLookupError:
+        return False
+    except PermissionError:
+        return True  # exists but owned by another user
+    return True
+
+
 def connect(name: str, extra_args: list[str] | None = None) -> dict[str, Any]:
     integration = get_integration(name)
     if integration is None:
@@ -35,6 +67,9 @@ def connect(name: str, extra_args: list[str] | None = None) -> dict[str, Any]:
 
     proc = integration.start(extra_args=extra_args)
     _RUNNING[name] = proc
+    # Persist the pid so a later `tether connect down` (a separate CLI process)
+    # can actually stop this child instead of orphaning it.
+    _write_pid(name, proc.pid)
 
     return {
         "status": "started",
@@ -45,14 +80,48 @@ def connect(name: str, extra_args: list[str] | None = None) -> dict[str, Any]:
     }
 
 
+def _signal_for(integration) -> int:
+    sig_name = getattr(integration, "stop_signal", "SIGTERM") if integration else "SIGTERM"
+    return getattr(signal, sig_name, signal.SIGTERM)
+
+
 def disconnect(name: str) -> dict[str, Any]:
+    integration = get_integration(name)
+
+    # 1) Same-process handle (rare — only if up + down ran in one process).
     proc = _RUNNING.pop(name, None)
     if proc is not None and proc.poll() is None:
-        proc.terminate()
-        proc.wait(timeout=10)
+        proc.send_signal(_signal_for(integration))
+        try:
+            proc.wait(timeout=10)
+        except subprocess.TimeoutExpired:
+            proc.kill()
+            proc.wait(timeout=5)
+        _pid_file(name).unlink(missing_ok=True)
         return {"status": "stopped", "name": name, "pid": proc.pid}
 
-    integration = get_integration(name)
+    # 2) Cross-invocation: signal the pid we persisted at connect time.
+    pid = _read_pid(name)
+    if pid is not None and _pid_alive(pid):
+        sig = _signal_for(integration)
+        try:
+            os.kill(pid, sig)
+            # Wait for it to actually exit; escalate to SIGKILL if it won't.
+            for _ in range(10):
+                if not _pid_alive(pid):
+                    break
+                time.sleep(1)
+            else:
+                os.kill(pid, signal.SIGKILL)
+        except ProcessLookupError:
+            pass
+        _pid_file(name).unlink(missing_ok=True)
+        return {"status": "stopped", "name": name, "pid": pid}
+
+    # 3) Stale pid file (process already gone) — clean it up.
+    _pid_file(name).unlink(missing_ok=True)
+
+    # 4) Something is still answering on the port but we didn't start it.
     if integration and integration.health_check():
         return {"status": "external_still_running", "name": name}
 
diff --git a/src/tether/integrations/registry.py b/src/tether/integrations/registry.py
index 6d1c882..cdb8fb3 100644
--- a/src/tether/integrations/registry.py
+++ b/src/tether/integrations/registry.py
@@ -1,26 +1,45 @@
 """Integration registry — known external tools tether can connect to."""
 from __future__ import annotations
 
-import importlib
+import importlib.util
 import logging
-import shutil
+import os
 import subprocess
 import sys
 import time
 from dataclasses import dataclass, field
-from typing import Any
+from pathlib import Path
 
 import requests
 
 logger = logging.getLogger(__name__)
 
 
+def state_dir() -> Path:
+    """Per-user dir for integration runtime state (pid + log files).
+
+    Honors TETHER_HOME so it agrees with the rest of the CLI's config home.
+    """
+    home = Path(os.environ.get("TETHER_HOME", Path.home() / ".tether"))
+    d = home / "integrations"
+    d.mkdir(parents=True, exist_ok=True)
+    return d
+
+
 @dataclass(frozen=True)
 class Integration:
     name: str
     description: str
     pip_package: str
     pip_extras: str = ""
+    # Module name to import for the installed-check. Defaults to the pip
+    # package with '-' → '_'; set explicitly when they differ (the common
+    # case for any package whose import name isn't its pip name).
+    import_name: str = ""
+    # Optional version constraint applied at install time (e.g. ">=1.2,<2").
+    # Empty = unpinned (a future release of the integration can change
+    # behavior silently — set this for anything load-bearing).
+    pip_version_spec: str = ""
     health_url: str = "http://localhost:8000/healthz"
     start_command: list[str] = field(default_factory=list)
     stop_signal: str = "SIGTERM"
@@ -31,24 +50,38 @@ class Integration:
 
     @property
     def pip_spec(self) -> str:
+        spec = self.pip_package
         if self.pip_extras:
-            return f"{self.pip_package}[{self.pip_extras}]"
-        return self.pip_package
+            spec = f"{spec}[{self.pip_extras}]"
+        if self.pip_version_spec:
+            spec = f"{spec}{self.pip_version_spec}"
+        return spec
+
+    @property
+    def _import_name(self) -> str:
+        return self.import_name or self.pip_package.replace("-", "_")
+
+    @property
+    def log_file(self) -> Path:
+        return state_dir() / f"{self.name}.log"
 
     def is_installed(self) -> bool:
+        # find_spec resolves the module WITHOUT importing it — so checking
+        # "is it installed?" doesn't run the target package's import-time side
+        # effects (rtsm[gpu] can init CUDA on import).
         try:
-            importlib.import_module(self.pip_package.replace("-", "_").split("[")[0])
-            return True
-        except ImportError:
+            return importlib.util.find_spec(self._import_name) is not None
+        except (ImportError, ValueError):
             return False
 
     def install(self) -> None:
-        logger.info("Installing %s...", self.pip_spec)
-        cmd = [
-            sys.executable, "-m", "pip", "install", self.pip_spec,
-            "--extra-index-url", "https://download.pytorch.org/whl/cu128",
-            "-q",
-        ]
+        logger.info("Installing %s into %s ...", self.pip_spec, sys.executable)
+        cmd = [sys.executable, "-m", "pip", "install", self.pip_spec]
+        # The PyTorch CUDA wheel index is only correct on Linux + CUDA. On
+        # macOS, Jetson (JetPack wheels), and CPU boxes it's useless or wrong,
+        # so don't force it there.
+        if sys.platform.startswith("linux") and "cu" in self.pip_extras.lower():
+            cmd += ["--extra-index-url", "https://download.pytorch.org/whl/cu128"]
         subprocess.check_call(cmd)
         logger.info("Installed %s", self.pip_spec)
 
@@ -63,23 +96,44 @@ def start(self, extra_args: list[str] | None = None) -> subprocess.Popen:
         cmd = list(self.start_command)
         if extra_args:
             cmd.extend(extra_args)
-        logger.info("Starting %s: %s", self.name, " ".join(cmd))
-        proc = subprocess.Popen(
-            cmd,
-            stdout=subprocess.PIPE,
-            stderr=subprocess.PIPE,
-        )
-        for _ in range(30):
-            if self.health_check(timeout=1.0):
-                logger.info("%s is healthy at %s", self.name, self.health_url)
-                return proc
-            time.sleep(1)
+        log_path = self.log_file
+        logger.info("Starting %s: %s (logs → %s)", self.name, " ".join(cmd), log_path)
+        # Redirect child output to a log file, NOT PIPE: an unread PIPE fills
+        # its 64KB buffer and the child blocks, looking like a hang. The log
+        # file also gives `connect` a real place to point users on failure.
+        log_fh = open(log_path, "wb")
+        proc = subprocess.Popen(cmd, stdout=log_fh, stderr=subprocess.STDOUT)
+        try:
+            for _ in range(30):
+                # Fail fast if the child died on launch (bad flag, missing
+                # model) instead of waiting the full 30s for a health check
+                # that can never pass.
+                if proc.poll() is not None:
+                    tail = self._log_tail(log_path)
+                    raise RuntimeError(
+                        f"{self.name} exited immediately (code {proc.returncode}). "
+                        f"Last log lines:\n{tail}\nFull log: {log_path}"
+                    )
+                if self.health_check(timeout=1.0):
+                    logger.info("%s is healthy at %s", self.name, self.health_url)
+                    return proc
+                time.sleep(1)
+        finally:
+            log_fh.close()
         proc.terminate()
+        tail = self._log_tail(log_path)
         raise RuntimeError(
             f"{self.name} failed to become healthy at {self.health_url} "
-            f"within 30s. Check logs with: {' '.join(cmd)}"
+            f"within 30s. Last log lines:\n{tail}\nFull log: {log_path}"
         )
 
+    @staticmethod
+    def _log_tail(path: Path, n: int = 20) -> str:
+        try:
+            return "\n".join(path.read_text(errors="replace").splitlines()[-n:])
+        except OSError:
+            return "(log unavailable)"
+
 
 RTSM = Integration(
     name="rtsm",
diff --git a/tests/test_integrations_lifecycle.py b/tests/test_integrations_lifecycle.py
new file mode 100644
index 0000000..d28aea1
--- /dev/null
+++ b/tests/test_integrations_lifecycle.py
@@ -0,0 +1,95 @@
+"""Integration connect/disconnect lifecycle.
+
+Pins the fix for `tether connect down` being a structural no-op (the running
+handle lived in a module dict that died with the CLI process) by persisting the
+pid, plus the registry hardening (find_spec install-check, fail-fast start,
+version-pinned spec).
+"""
+from __future__ import annotations
+
+import subprocess
+import sys
+
+import pytest
+
+from tether.integrations import connector
+from tether.integrations.registry import Integration
+
+
+@pytest.fixture(autouse=True)
+def _isolated_home(tmp_path, monkeypatch):
+    monkeypatch.setenv("TETHER_HOME", str(tmp_path))
+    connector._RUNNING.clear()
+    yield
+
+
+def test_disconnect_stops_pid_from_a_separate_process():
+    """The pid persisted at connect time lets a fresh process actually stop it."""
+    proc = subprocess.Popen([sys.executable, "-c", "import time; time.sleep(30)"])
+    try:
+        connector._write_pid("rtsm", proc.pid)
+        connector._RUNNING.clear()  # simulate: `connect down` is a new CLI process
+        assert connector._pid_alive(proc.pid)
+
+        result = connector.disconnect("rtsm")
+
+        assert result["status"] == "stopped"
+        assert result["pid"] == proc.pid
+        proc.wait(timeout=5)
+        assert not connector._pid_alive(proc.pid)
+        assert not connector._pid_file("rtsm").exists()
+    finally:
+        if proc.poll() is None:
+            proc.kill()
+            proc.wait(timeout=5)
+
+
+def test_disconnect_cleans_stale_pid_file():
+    dead = subprocess.Popen([sys.executable, "-c", "pass"])
+    dead.wait()
+    connector._write_pid("rtsm", dead.pid)
+    connector._RUNNING.clear()
+
+    result = connector.disconnect("rtsm")
+
+    assert result["status"] in ("not_running", "external_still_running")
+    assert not connector._pid_file("rtsm").exists()
+
+
+def test_is_installed_uses_find_spec_no_import():
+    present = Integration(name="a", description="", pip_package="os")
+    assert present.is_installed() is True
+    absent = Integration(name="b", description="", pip_package="totally-not-real-pkg-xyz")
+    assert absent.is_installed() is False
+
+
+def test_is_installed_honors_explicit_import_name():
+    # pip name != import name: declare it.
+    i = Integration(name="c", description="", pip_package="pip-thing", import_name="json")
+    assert i.is_installed() is True
+
+
+def test_start_fails_fast_on_immediate_exit():
+    integ = Integration(
+        name="boom",
+        description="",
+        pip_package="x",
+        start_command=[sys.executable, "-c", "import sys; sys.exit(3)"],
+        health_url="http://localhost:59999/healthz",
+    )
+    with pytest.raises(RuntimeError) as ei:
+        integ.start()
+    assert "exited immediately" in str(ei.value)
+    assert integ.log_file.exists()
+
+
+def test_pip_spec_includes_extras_and_version():
+    i = Integration(
+        name="z", description="", pip_package="rtsm",
+        pip_extras="gpu", pip_version_spec=">=1.0,<2",
+    )
+    assert i.pip_spec == "rtsm[gpu]>=1.0,<2"
+
+
+def test_connect_disconnect_exported():
+    from tether.integrations import connect, disconnect  # noqa: F401