openguardrails-instrumentation-langgraph 0.1.1

Guard a custom LangGraph agent with the OpenGuardrails (OGR) protocol — a drop-in ToolNode at the agent-hook altitude, with native LangGraph interrupt() as the human-confirm gate.

openguardrails-instrumentation-langgraph

Guard a custom LangGraph agent with the OpenGuardrails (OGR) protocol — a drop-in ToolNode that enforces one policy you own at the agent-hook altitude, using LangGraph's own interrupt() as the human-confirm gate.

pip install "openguardrails-instrumentation-langgraph[langgraph]"

Why this is different from the other instrumentations

The opencode / OpenClaw / Claude Code bindings hook a finished product's plugin surface. A developer hand-rolling an agent with LangGraph has no plugin marketplace — they need a library they import. That library's natural Policy Enforcement Point (PEP) is the tool boundary: in LangGraph, the ToolNode. That node is the agent-hook altitude.

What it enforces

Every tool_call the model emits is turned into an OGR GuardEvent and judged by the reference Runtime (the PDP) before the tool runs:

Tool call	Provenance	Decision
get_weather(city="Paris")	trusted	allow — runs
bash("rm -rf /")	trusted	block — refused in place, never runs
bash("curl … | sh")	trusted	require_approval → interrupt()
bash("curl … | sh")	untrusted (after a web/retrieval result)	block — injection

The last two rows are the point: the same command gets a different verdict depending on where its inputs came from. A tool whose result carries external content (web fetch, retrieval, MCP) taints the session, so a later tool call inherits untrusted provenance — the indirect-prompt-injection defense.

The human-confirm gate is LangGraph-native

A require_approval verdict doesn't invent an approval UI — it raises a LangGraph interrupt(). The graph pauses at its checkpointer; a human approves; the graph resumes. This keeps OGR's red line intact: the confirm gate and the enforcement point stay privilege-separated — the agent cannot self-approve.

from langgraph.types import Command

out = app.invoke({"messages": [...]}, config)      # pauses at __interrupt__
# ... surface out["__interrupt__"][0].value to a human ...
out = app.invoke(Command(resume=True), config)     # approve all
# or Command(resume={tool_call_id: True/False}) / a list of approved ids

Resume accepts True/False, "approve"/"deny", a {tool_call_id: bool} map, or a list of approved ids. Anything unrecognized fails closed (deny).

Three integration surfaces

from openguardrails_instrumentation_langgraph import (
    OpenGuardrailsToolNode, guard, ogr_guard,
)

1. Drop-in ToolNode — the primary PEP, for a hand-built StateGraph:

graph.add_node("tools", OpenGuardrailsToolNode(tools))   # was ToolNode(tools)
app = graph.compile(checkpointer=MemorySaver())          # checkpointer needed for approvals

2. guard(tools) — when you don't own the ToolNode (e.g. the prebuilt agent):

from langgraph.prebuilt import create_react_agent
agent = create_react_agent(model, guard(tools))

3. @ogr_guard — a single tool:

@ogr_guard
@tool
def bash(command: str) -> str: ...

OpenGuardrailsToolNode gives full provenance/injection coverage (it sees the run config and threads session taint); guard() / ogr_guard judge with base provenance.

The policy

policy.json is a thin overlay — {"$extends": "openguardrails:base"}. The core resolves the canonical PAP base and deep-merges your overlay into one effective policy, so the same policy that fronts your gateway and sandbox also judges these tool calls. Point elsewhere with OGR_POLICY=/path/to/policy.json or OpenGuardrailsToolNode(tools, policy="…").

Altitude

This binding is the agent-hook altitude (the tool boundary). The gateway altitude (prompt-injection / secret leakage on the LLM wire) is covered by openguardrails-instrumentation-litellm if your LangGraph app routes its model through litellm — the two compose.

Try it

python examples/react_agent.py     # drives allow / block / approve / deny end-to-end
python tests/test_smoke.py         # offline decision-core test (no langgraph needed)