Installation

Requirements

• Python 3.9 or newer.
• A platform supported by sentence-transformers and PyTorch. The semantic firewall is the only stage that needs them; the deterministic stages do not, but the base install includes them.

Install from a clone

git clone https://github.com/CharafeddineMechalikh/GATE.git
cd GATE
python -m venv .venv
# Linux / macOS
source .venv/bin/activate
# Windows (Git Bash)
source .venv/Scripts/activate

pip install -e ".[all,dev]"

all installs the four framework extras. dev adds pytest and the linting tools used by the test suite. To install only the frameworks you use, pick one or more extras:

pip install -e ".[langgraph,dev]"

Extras

The audit extra is intentionally empty. The encrypted logger uses the standard library for the local and Vault backends. The AWS KMS backend requires boto3, which is not installed by default:

pip install boto3

Framework versions

The integration suite is verified against a single environment containing:

Those versions are not pinned by this package. Framework internals change, and the adapters depend on a small number of stable dispatch points (BaseTool._run/_arun, FunctionTool.on_invoke_tool, LangGraph ToolNode.tools_by_name). If a newer framework release changes one of those points, the integration tests fail loudly rather than silently downgrading enforcement.

The embedding model

DefaultSemanticFirewall constructs a sentence-transformers model named by gateConfig.embedding_model_name, which defaults to all-MiniLM-L6-v2. The first semantic check downloads the model weights (roughly 90 MB) unless they are already cached.

To avoid a download on first use in an offline deployment, preload the model once in the target environment, or set embedding_model_name to a model already present in the local Hugging Face cache. The model is loaded when the firewall is constructed, which happens when a gateMonitor is built with the semantic stage enabled.

The deterministic stages (privilege, dfa, contracts) and the D-CBF stage do not need the embedding model. Configure enabled_checks={"privilege", "dfa", "contracts", "dcbf"} to run without it.

Verify the installation

import gate
from gate import gateConfig, AgentProfile

print(gate.__version__)  # 1.1.1

config = gateConfig(
    agent_profile=AgentProfile(
        agent_id="smoke-test",
        role="tester",
        allowed_tools={"echo"},
    ),
    enabled_checks={"privilege"},
)
print(config.agent_profile.agent_id)

To verify the adapters, run the integration suite:

pytest tests/test_framework_adapters.py -v

These tests build real framework objects and exercise native dispatch. They do not call a model, so they need no API key and no network access once the framework packages are installed.

Windows notes

The adapters enforce invocation timeouts. On POSIX systems in the main thread, a timeout interrupts the call with SIGALRM. On Windows, Python cannot interrupt the running call, so the adapter stops waiting at the deadline, records the timeout, and calls ToolEnforcement.abort() to deny any further tool calls from that run. The worker thread itself is not forcibly killed. Do not rely on a timeout to undo a tool body that has already started executing on Windows.

Upgrading

The 0.3.0 release changed the monitor API to async (validate_action, validate_result) and introduced the shared enforcement layer. Code written against 0.1.x that called monitor methods synchronously, or that assumed adapters only observed calls, must be updated. See the changelog.