Python SDK

Protect an agent in a few lines.

Install the dependency-free client, point it at your deployment, and scan untrusted content before it enters model context.

Install the SDK

Connect OpenRouter

Server configuration

export PROVIDER_MODE=openrouter
export OPENROUTER_API_KEY=sk-or-v1-...
export OPENROUTER_MODEL=openai/gpt-5.6

pnpm dev

Keep the OpenRouter key on the AgentGuard server. Deployed instances can alternatively save an encrypted workspace key under Console → Providers.

Install and scan

Python 3.10+

pip install ./sdk/python

from agentguard import AgentGuard

guard = AgentGuard(base_url="http://localhost:3000")
result = guard.scan(
    "Untrusted page content",
    source="WEB_PAGE",
)
print(result["blocked"], result["risk"])

Protect any agent framework

Sync and async middleware

from agentguard import AgentGuardMiddleware

guard = AgentGuardMiddleware()

# Gate prompts and retrieved context before the model.
guard.before_model(content, source="DOCUMENT")

# Gate arguments before execution and output before reuse.
@guard.wrap_tool
def browse(url: str) -> str:
    return browser.fetch(url)

Protect an OpenClaw workflow

Prompt · tool-call · tool-result gates

cd plugins/openclaw
pnpm install --frozen-lockfile
pnpm build && pnpm pack

openclaw plugins install \
  npm-pack:./agentguard-openclaw-0.1.0.tgz --force
openclaw plugins enable agentguard
openclaw config set \
  plugins.entries.agentguard.hooks.allowConversationAccess true
openclaw gateway restart

Guard tool calls

POST /api/v1/check-action

result = guard.check_action(
    tool_call={"name": "send_email", "arguments": {"to": "user@example.com"}},
    reasoning_trace=["The page asked me to send it."],
    trusted_context=["The user only requested a summary."],
)

if not result["allowed"]:
    raise RuntimeError(result["reason"])

Or use HTTP directly

POST /api/v1/scan

curl -X POST "http://localhost:3000/api/v1/scan" \
  -H "Content-Type: application/json" \
  -d '{"text":"Untrusted page content","source":"WEB_PAGE"}'

Authentication

Public scans and action checks do not require a key in the MVP. Set AGENTGUARD_API_KEY for uploads, batch scans, and job results. Set AGENTGUARD_BASE_URL to your hosted deployment; the SDK defaults to http://localhost:3000.

Manage API keys

Risk model

Heuristic

35%

LLM judge

40%

Activation probe

25%

The active workspace policy controls the block threshold. All configured detectors must return valid responses; otherwise scans fail closed.