feat(desktop): allow specifying a remote letta server in the letta desktop (#3709)

letta/orm/sqlite_functions.py

@@ -152,7 +152,7 @@ def register_functions(dbapi_connection, connection_record):
             if is_aiosqlite_connection:
                 # For aiosqlite connections, we cannot use async operations in sync event handlers
                 # The extension will need to be loaded per-connection when actually used
-                logger.info("Detected aiosqlite connection - sqlite-vec will be loaded per-query")
+                logger.debug("Detected aiosqlite connection - sqlite-vec will be loaded per-query")
             else:
                 # For sync connections
                 # dbapi_connection.enable_load_extension(True)
@@ -173,7 +173,7 @@ def register_functions(dbapi_connection, connection_record):
                 raw_conn = getattr(actual_connection, "_connection", actual_connection)
                 if hasattr(raw_conn, "create_function"):
                     raw_conn.create_function("cosine_distance", 2, cosine_distance)
-                    logger.info("Successfully registered cosine_distance for aiosqlite")
+                    logger.debug("Successfully registered cosine_distance for aiosqlite")
             else:
                 dbapi_connection.create_function("cosine_distance", 2, cosine_distance)
                 logger.info("Successfully registered cosine_distance for sync connection")

letta/server/db.py

@@ -226,7 +226,12 @@ class DatabaseRegistry:
 
     def _build_sqlalchemy_engine_args(self, *, is_async: bool) -> dict:
         """Prepare keyword arguments for create_engine / create_async_engine."""
-        use_null_pool = settings.disable_sqlalchemy_pooling
+        # For async SQLite, always use NullPool to avoid cleanup issues during cancellation
+        if is_async and settings.database_engine is DatabaseChoice.SQLITE:
+            use_null_pool = True
+            logger.info("Forcing NullPool for async SQLite to avoid cancellation cleanup issues")
+        else:
+            use_null_pool = settings.disable_sqlalchemy_pooling
 
         if use_null_pool:
             logger.info("Disabling pooling on SqlAlchemy")
@@ -262,7 +267,8 @@ class DatabaseRegistry:
                     }
                 )
 
-        elif is_async:
+        elif is_async and settings.database_engine is DatabaseChoice.POSTGRES:
+            # Invalid for SQLite, results in [0] TypeError: 'prepared_statement_name_func' is an invalid keyword argument for Connection()
             # For asyncpg, statement_cache_size should be in connect_args
             base_args.update(
                 {

letta/services/tool_sandbox/local_sandbox.py

@@ -120,6 +120,10 @@ class AsyncToolSandboxLocal(AsyncToolSandboxBase):
             else:
                 # If not using venv, use whatever Python we are running on
                 python_executable = sys.executable
+                # For embedded/desktop environments, preserve Python paths
+                # This ensures the subprocess can find bundled modules
+                if "PYTHONPATH" in os.environ:
+                    exec_env["PYTHONPATH"] = os.environ["PYTHONPATH"]
 
             # handle unwanted terminal behavior
             exec_env.update(
@@ -202,7 +206,7 @@ class AsyncToolSandboxLocal(AsyncToolSandboxBase):
                     except asyncio.TimeoutError:
                         process.kill()
 
-                raise TimeoutError(f"Executing tool {self.tool_name} timed out after 60 seconds.")
+                raise TimeoutError(f"Executing tool {self.tool_name} timed out after {tool_settings.tool_sandbox_timeout} seconds.")
 
             stderr = stderr_bytes.decode("utf-8") if stderr_bytes else ""
             log_event(name="finish subprocess")

poetry.lock

@@ -8719,7 +8719,7 @@ cffi = ["cffi (>=1.11)"]
 all = ["autoflake", "black", "docker", "fastapi", "google-cloud-profiler", "granian", "isort", "langchain", "langchain-community", "locust", "pexpect", "pg8000", "pgvector", "pinecone", "pre-commit", "psycopg2", "psycopg2-binary", "pyright", "pytest-asyncio", "pytest-order", "redis", "turbopuffer", "uvicorn", "uvloop", "wikipedia"]
 bedrock = ["aioboto3", "boto3"]
 cloud-tool-sandbox = ["e2b-code-interpreter", "modal"]
-desktop = ["docker", "fastapi", "langchain", "langchain-community", "locust", "pyright", "sqlite-vec", "uvicorn", "wikipedia"]
+desktop = ["docker", "fastapi", "langchain", "langchain-community", "locust", "pgvector", "pyright", "sqlite-vec", "uvicorn", "wikipedia"]
 dev = ["autoflake", "black", "isort", "locust", "pexpect", "pre-commit", "pyright", "pytest-asyncio", "pytest-order"]
 experimental = ["google-cloud-profiler", "granian", "uvloop"]
 external-tools = ["docker", "firecrawl-py", "langchain", "langchain-community", "wikipedia"]
@@ -8733,4 +8733,4 @@ tests = ["wikipedia"]
 [metadata]
 lock-version = "2.0"
 python-versions = "<3.14,>=3.11"
-content-hash = "349c668d93463b0f4e4526311fd32aad465e139148dd597d1411c17ffcd43c89"
+content-hash = "157ecc4bb05492f7b8c7758e2af195d635e42ba7e8cb17e6be276844b1dbd4b4"

pyproject.toml

@@ -120,7 +120,7 @@ external-tools = ["docker", "langchain", "wikipedia", "langchain-community", "fi
 tests = ["wikipedia"]
 bedrock = ["boto3", "aioboto3"]
 google = ["google-genai"]
-desktop = ["pyright", "websockets", "fastapi", "uvicorn", "docker", "langchain", "wikipedia", "langchain-community", "locust", "sqlite-vec"]
+desktop = ["pyright", "websockets", "fastapi", "uvicorn", "docker", "langchain", "wikipedia", "langchain-community", "locust", "sqlite-vec", "pgvector"]
 all = ["pgvector", "turbopuffer", "pg8000", "psycopg2-binary", "psycopg2", "pytest", "pytest-asyncio", "pexpect", "black", "pre-commit", "pyright", "pytest-order", "autoflake", "isort", "websockets", "fastapi", "uvicorn", "docker", "langchain", "wikipedia", "langchain-community", "locust", "uvloop", "granian", "redis", "pinecone", "google-cloud-profiler"]
 
 [tool.poetry.group.dev.dependencies]