Skip to content
Developer documentation

Put the decision before the action.

Onboard the shared execution boundary once, then route eligible tool and workflow interactions through it. This is the public contract-level guide; customer configuration, credentials, and policy internals remain private.

Choose a connection

Gateway: configure supported MCP, API, GraphQL and A2A calls through your assigned sidecar gateway, without adding AgenticDome SDK hooks.

SDKs and plugins: add checks at supported application execution hooks for internal tools and deeper framework context. Combine both paths under shared tenant policy.

Test allowed and denied actions on the paths you connect. Use runtime coverage to review protection and identify gaps.

Gateway quickstart

  1. In your signed-in Action Gateway workspace, select an assigned sidecar and copy its gateway connection settings and Runtime key.
  2. Your security or AI team configures permitted destination domains and trusted caller identities.
  3. Choose the native client example for your protocol or framework. Supply your existing service URL, the protocol request and the appropriate caller identity.
  4. Test a permitted and a denied request, then review the recorded decision.

The gateway URL comes from your assigned sidecar. The destination URL is the service you already use. Individual destination registration is not required; the destination must be within your configured boundary.

Examples cover Docker Agent, AWS Strands, CrewAI and LangGraph. MCP clients retain their protocol message structure; AgenticDome adds the policy evaluation and decision evidence around the routed request.

MCP input schemas: the updated Gateway obtains the accessible tool catalogue using the same server access and validates arguments against published inputSchema before forwarding. It caches compiled schemas locally. A valid argument structure still needs identity and action-policy approval.

Discovery records reported schema hashes and changes. Optionally pin a sensitive tool’s approved input contract in Tool & MCP Registry. Acknowledging a discovery change does not approve that contract for execution.

Explore the gateway connection · Compare native clients

Docker Agent local tools

Use the Gateway for remote service requests. Add Docker’s mandatory tool_guard when the agent also proposes shell, file or local-tool actions. The adapter sends the proposal to the sidecar; Docker executes it only after an allow decision, including auto-approved tool calls.

  1. In Action Gateway → Onboarding, choose your assigned sidecar and open Docker native tool setup.
  2. Install the guard adapter from a published SDK release containing agenticdome-docker-guard. Use a native Docker Agent runtime supporting mandatory tool_guard and global hooks.d configuration.
  3. Download the generated 50-agenticdome.yaml to the runtime’s configured hooks.d directory. It contains the actual assigned Gateway URL and tenant. Load the Runtime API key and approved identity settings through your workload’s secret manager.
  4. Run agenticdome-docker-guard --check, confirm Docker loaded the hook, and test allowed and denied proposals in a disposable workspace.

Your application code stays the same. Keep the adapter and hook outside agent-writable mounts and retain Docker’s sandbox and access controls. Guard decisions are pre-execution policy evidence; review separate action and outcome evidence to establish execution.

Explore both Docker pathways · Plan the setup with Abi

Caller identity

Your Runtime key connects the application to the tenant. Verified caller identity tells policy which workload or agent is acting, and which user it represents when authority is delegated.

Bring an existing provider such as Microsoft Entra ID or Okta, or enroll an approved deployment for AgenticDome-issued agent identities. The signed-in Identity workspace supplies the settings for your chosen approach.

Compare the two identity paths

SDK quickstart · Python

import os
from agenticdome_sdk.client import AgenticDomeClient

client = AgenticDomeClient(
    api_base=os.environ["AGENTICDOME_API_BASE"],
    api_key=os.environ["AGENTICDOME_API_KEY"],
    tenant_id=os.environ["AGENTICDOME_TENANT_ID"],
)

decision = client.guardrail_validate(
    text="Agent requests a refund tool call.",
    agent_id="support-agent",
    direction="outbound",
    platform="python",
    tool_name="payments.refund.create",
    tool_args={"amount": 2500, "currency": "USD"},
)

if decision.get("verdict") == "BLOCKED":
    raise PermissionError(decision.get("reason", "Action blocked"))

Copy the real API base, Runtime key and tenant ID from your account's assigned runtime settings into server-side configuration. The example requests a decision; use the supported framework integration to enforce decisions, carry identity and record outcomes at execution.

Action-decision contract

The Action Firewall returns a decision such as ALLOWED, BLOCKED, or REDACTED, together with decision context available to the integration. The application remains responsible for enforcing that result before the side effect occurs.

Compatibility paths

Public implementation paths cover Python and TypeScript applications, REST, MCP/tool boundaries, OpenClaw workspaces, Microsoft Copilot Studio and AI Foundry patterns, AWS Bedrock agents, and common agent frameworks. Platform names describe integration patterns unless a page explicitly identifies a native package or certified connector.

Plan your approach with Abi

Describe your workflow, stack and constraints. Abi helps compare supported approaches and prepares a deployment blueprint for review. Existing customers can open Ask Abi in their signed-in workspace.

AgenticDome Copilot guides the supported onboarding and verification steps once you are ready to connect.

Meet Abi · Explore guided onboarding

Runtime performance

AgenticDome is engineered for inline runtime use. Warming and deterministic policy paths are designed to keep AgenticDome processing to a few milliseconds for eligible decisions. Do not treat component processing time as total request overhead: network RTT, proxy traversal, payload size, retries, and optional LLM-assisted branches can add materially more.

Co-locate the assigned runtime with the protected workload where possible, reuse connections, choose bounded timeouts, and run the performance harness against representative allowed, blocked, and unavailable paths before production.

Timeout and error semantics

Set explicit timeouts and select fail-closed or fail-open behavior based on the action’s risk. Test network failure, malformed decisions, retries, and duplicate submissions. Production-sensitive actions should not silently execute when their required decision is unavailable.

See the Trust Center for public boundary disclosures or book a deployment review for your exact architecture.