agent-guard-python

agent-guard-python

Python bindings for execution control at the agent side-effect boundary.

This package provides Python bindings for agent-guard, giving Python hosts a pre-execution decision layer before agent tool calls turn into shell commands or other side effects.

---

🚀 Quick Start (Python Adapters)

Status: The Python wrapper layer is a beta adapter surface. The clearest current proof point is still shell-first execution control, and Node remains the most mature integration path in the repository.

Integrate agent-guard into your existing LangChain tools with a policy gate in front of the original tool:

from agent_guard import Guard, wrap_langchain_tool

# 1. Initialize the Guard with your security policy
guard = Guard.from_yaml_file("policy.yaml")

# 2. Secure your existing tools
bash_tool = ShellTool() # Your original tool
secured_tool = wrap_langchain_tool(guard, bash_tool, agent_id="researcher")

# 3. Use the tool as normal - agent-guard handles the rest!
secured_tool.run("ls -la")

OpenAI-style handler wrapping is also available:

from agent_guard import Guard, wrap_openai_tool, AgentGuardDeniedError

guard = Guard.from_yaml_file("policy.yaml")

guarded_handler = wrap_openai_tool(
    guard,
    lambda input_data: {"ok": True, "query": input_data["query"]},
    tool="web_search",
    mode="check",
    trust_level="trusted",
)

try:
    print(guarded_handler({"query": "agent-guard"}))
except AgentGuardDeniedError as error:
    print("blocked", error.code)

---

✨ Features

• 🛡️ Pre-execution policy decisions: Put allow/deny/ask checks in front of Python tool handlers.
• 💻 Shell-first execution control: The strongest current execution path is still shell / Bash style tooling.
• ⚠️ Typed adapter errors: Distinct deny, ask-required, and execution failure exceptions.
• 📜 Signed receipts: Optional cryptographic proof of execution when you need deeper verification.
• 🔏 Signed policy loading: Optional detached-signature verification for policy.yaml.
• 📊 Auditing support: JSONL logs and metrics integration for operator-visible outcomes.

Current boundary note:

• explicit check mode remains a policy-only gate; in auto mode, non-shell tools use Guard.run(). A Handoff closes its one-shot lifecycle before returning the original handler result, while Guard-owned WriteFile or mutating HTTP outcomes may execute without invoking that handler
• the exact tool ID bash is the owned shell path; shell-like custom IDs such as shell, terminal, sh, cmd, or powershell fail closed in auto with UnsupportedShellAlias until explicitly mapped or assigned a mode
• Python is an active adapter surface, but still below the current Node path in maturity
• Guard.execute() / Guard.run() use the SDK default sandbox selection, or accept an explicit backend= keyword ("none", "linux-seccomp", "linux-landlock", "macos-seatbelt", "windows-job-object", "windows-appcontainer"); a backend that is not compiled into the module or not functional on the host resolves truthfully to "none", and an unknown name raises
• to get real isolation through backend=, build the module with the matching feature forwarded, e.g. maturin develop --features extension-module,seccomp (requires libseccomp on Linux); the default build carries no sandbox feature
• if the default sandbox diagnosis falls back to NoopSandbox, the policy gate still runs, but OS-level isolation is not equivalent
• unknown, expired, or duplicate handoff report IDs raise GuardError; a successful host action whose report fails surfaces AgentGuardExecutionError with the host result, request ID, attempted handoff result, and an explicit do-not-retry marker rather than silently returning success

---

🔧 Installation

The latest published package is 0.2.6:

python -m pip install agent-guard-python==0.2.6

To test the current 0.2.8 source, install it from a repository checkout (requires Python and a Rust toolchain):

python -m pip install ./crates/agent-guard-python

The distribution name is agent-guard-python; the import name remains agent_guard:

from agent_guard import Guard

For local development in this repository:

cd crates/agent-guard-python
maturin develop --features extension-module
pytest tests -v

---

📺 Demos

Check the examples/ directory for full usage scenarios:

• demo_langchain.py: Comprehensive 3-line integration demo.
• provenance_receipt.py: Cryptographic verification example.