One init(), 9 frameworks.

Supported frameworks

• Anthropic SDK (Python + JS)
• OpenAI SDK (Python + JS)
• LangChain + LangGraph (Python + JS)
• CrewAI
• LlamaIndex
• Mistral AI
• Google Gemini / Vertex AI
• AWS Bedrock (boto3)
• smolagents

If your framework isn't on this list, the LLM egress proxy mode works for anything that respects OPENAI_BASE_URL / ANTHROPIC_BASE_URL / equivalent.

Python SDK

pip install agentguard

# in your agent entrypoint:
import agentguard
agentguard.init(
    api_key="aeg_xxx",
    gateway_url="https://gateway.aegis.dev",  # or your self-hosted URL
    agent_id="my-agent",                       # optional but recommended
)

After init():

• Anthropic / OpenAI / Mistral / Gemini / Cohere clients are monkey-patched at import time. Every call goes through the gateway.
• LangChain BaseTool.run is patched — tool calls land as audit rows.
• CrewAI agent handoffs are intercepted for the cross-agent / collusion detector.
• Exceptions on blocked calls are AgentGuardBlockedError, with the policy id + reason.

JavaScript / TypeScript SDK

npm install @agentguard/sdk

// in your entrypoint:
import { init } from '@agentguard/sdk';
init({
  apiKey:     process.env.AEGIS_API_KEY,
  gatewayUrl: 'https://gateway.aegis.dev',
  agentId:    'my-agent',
  verbose:    false,  // set true (or AGENTGUARD_VERBOSE=1) for cold-start logs
});

Same auto-instrumentation surface as Python.

Configuration reference

Option	Required	Default	Notes
apiKey	yes	—	Org-scoped key from /signup or the cockpit Settings page.
gatewayUrl	yes	—	Hosted: https://gateway.aegis.dev. Self-host: http://localhost:8080.
agentId	no	auto	If unset, derived from the host + process. Set explicitly for multi-agent apps.
blocking	no	true	false = observe-only (no enforcement, just audit).
timeoutMs	no	5000	Gateway call deadline.
verbose	no	false	Cold-start "auto-patched: [...]" log line.

Manual mode (skip auto-instrumentation)

If you want explicit checkpoints instead of monkey-patching:

from agentguard import AgentGuard
guard = AgentGuard(api_key="aeg_xxx", gateway_url="...", agent_id="my-agent")

verdict = guard.check_tool_call(tool="db_query", arguments={"sql": "SELECT 1"})
if verdict.decision == "block":
    raise RuntimeError(verdict.reason)

Zero-code-change mode (no SDK)

If you'd rather not install anything, point your existing SDK's base URL at the AEGIS proxy:

# Anthropic / OpenAI / Mistral / Gemini all support a base URL env var.
export OPENAI_BASE_URL=https://gateway.aegis.dev/openai/v1
export AEGIS_API_KEY=aeg_xxx
# your code is unchanged

AEGIS proxies the call to upstream, runs detectors on the response (tool calls, embedded prompts, anomalies), writes audit, returns. Every framework that respects *_BASE_URL works zero-code.

Next

• REST API reference
• Policy DSL + templates
• Self-host guide