OpenAI Agents SDK

The OpenAI Agents SDK has no hook files. You call AgentLint from a tool input guardrail in your own code, using OpenAIAgentsAdapter.evaluate_tool_call().

Install

pip install agentlint openai-agents
agentlint init          # writes agentlint.yml

agentlint setup openai writes no files; it prints the snippet below.

Attach AgentLint to each function tool with a tool input guardrail:

from agents import Agent, Runner
from agents.decorators import tool

from agentlint.adapters.openai_agents import OpenAIAgentsAdapter

agentlint = OpenAIAgentsAdapter()


@tool(tool_input_guardrails=[agentlint.tool_input_guardrail("Bash")])
def run_shell(command: str) -> str:
    """Run a shell command in the project."""
    ...


@tool(
    tool_input_guardrails=[
        agentlint.tool_input_guardrail("Write", arguments={"file_path": "path", "content": "text"})
    ]
)
def write_file(path: str, text: str) -> str:
    """Write a file in the project."""
    ...


agent = Agent(name="builder", tools=[run_shell, write_file])
result = Runner.run_sync(agent, "List the files in the repo")

tool_input_guardrail(tool_name, arguments=None, project_dir=None) takes what the tool does in AgentLint terms (Bash, Write or Edit) and, if your tool's argument names differ, a mapping from AgentLint's keys (command, file_path, content, old_string, new_string) to yours. On an ERROR the SDK skips the call and returns the AgentLint reasons to the model (ToolGuardrailFunctionOutput.reject_content); otherwise the call proceeds.

To write the guardrail yourself, call agentlint.evaluate_tool_call(...) inside your own @tool_input_guardrail function. Guardrail APIs belong to the SDK; see its guardrails documentation. as_guardrail() from earlier versions returned a plain dict the SDK can't use and is deprecated.

reject_content skips the tool call and returns the AgentLint reasons to the model instead. Guardrail APIs belong to the SDK; check its guardrails documentation for your installed version.

Verify

Call the adapter directly:

python -c 'from agentlint.adapters.openai_agents import OpenAIAgentsAdapter as A; print(A().evaluate_tool_call("Bash", {"command": "rm -rf /"}))'

tripwire_triggered should be True. agentlint status does not cover this integration; it only reports hook-based agents.

What is checked

evaluate_tool_call(tool_name, tool_input, project_dir=None) evaluates one call as a PreToolUse event with the project's agentlint.yml and returns:

{
    "tripwire_triggered": bool,  # True if any finding is an ERROR
    "violations": [...],  # rule_id, message, severity, file_path, line, suggestion, ...
    "blocked_count": int,
    "warning_count": int,
}

Map each tool to the Claude-style name and argument shape that AgentLint rules expect:

Your tool does	tool_name	tool_input
Runs a shell command	Bash	{"command": "..."}
Writes a file	Write	{"file_path": "...", "content": "..."}
Edits a file	Edit	{"file_path": "...", "content": "..."}

The adapter's own names shell, file_write and file_edit are translated to these; any other name matches no tool-specific rules.

Tool guardrails only run for function tools and local MCP server tools. Hosted tools and the SDK's built-in shell and apply-patch tools do not go through this pipeline, so AgentLint cannot see those calls.

Configuration notes

• Project directory: the project_dir argument, then AGENTLINT_PROJECT_DIR, then OPENAI_PROJECT_DIR, then the current directory.
• evaluate_tool_call loads config and rules on every call and does not keep session state, record events or upload to AgentChute.
• For a lower-level API, see the Python embedding example in Generic integration.

Uninstall

Remove the guardrail from your tools. agentlint uninstall openai is a no-op.

Troubleshooting

• Guardrail never trips: check that you pass Bash, Write or Edit as tool_name with the argument names above.
• ModuleNotFoundError: openai.agents: import from agents, not openai.agents.
• General issues: Diagnostics.