Skip to content
Action security for AI agents

Govern every AI-agent action. Prove every outcome.

AgenticDome sits between AI agents and your business systems. It discovers activity, checks connected actions before they take effect, and enforces your action policy.

Try the local demo Book a deployment review
The local demo needs no account or API key.
One approved action. Two outcomes.
Approved refund USD 4,000 to payee P-1042 on behalf of a verified staff member
Amount and payee match the approval.ALLOW
Amount changed to USD 40,000; the approval no longer matches.BLOCK
Illustrative scenario. No payment is executed.

One connected platform

From the first connection
to the next decision.

Start with one workflow. Connect identity, policy and evidence as your agents grow.

01 · Connect

Start with the connection that fits.

Route MCP, API, GraphQL and A2A calls through your sidecar Gateway. Use supported SDK hooks for application context, or add Docker Agent’s mandatory native guard for local tools through configuration. Combine the paths as your needs grow.

Explore connection options

Find your starting point with Abi.
Meet your deployment concierge

The action-authority gap

Valid access. Wrong action.

An agent can hold valid credentials and still be asked to do something it should not. Credentials show who is calling. They do not show whether this specific action should run.

IdentityWho is calling?The verified agent and, when delegated, the person it acts for.
ContextWhy, and in what workflow?The task, the step and who delegated it.
ActionWhat exactly will happen?The tool, the payee, the amount, the payload.
How it works

One decision before business impact.

Discover what is running, connect the boundary, decide against one policy, verify with a real action, and keep watching for change.

  1. 1 · DiscoverFind agents, MCP servers and models, and how each was found.
  2. 2 · ConnectPut the check on the execution path: gateway, SDK, or both.
  3. 3 · DecideApply allow or block decisions and supported data controls; fail closed when required checks cannot complete.
  4. 4 · VerifyProve the check runs with a real action and evidence.
  5. 5 · MonitorReview retained tool, model and protection changes.

Follow one refund through the layers

Select a layer to see what it checks, or run the action and watch where the decision happens. Illustrative simulation; no payment or API call is made.

Runtime Grid uses configured workload observations and governed network evidence to help find paths around a check. Coverage applies to execution paths connected through supported integrations.
Gateway and SDK: where the check sits The gateway protects requests travelling to services. The SDK protects actions inside the agent application. Both use the same AgenticDome policy and runtime services, so the choice is where the check sits, not which rules apply. MCP argument checks use the server’s published input schema before forwarding. For Docker Agent’s local shell, file and tool proposals, add its mandatory native guard through deployment configuration—without changing application code. Explore the gateway · Bring or enrol agent identities · Plan your deployment with Abi
Discovery

See what exists before you protect it.

Connect a source and AgenticDome finds agents, MCP servers and models. Each record shows how it was found, whether it is actually used, and who called it. Then choose what to govern.

See how discovery works
Sample data
NameFound byIn useProtection
refund-agentSDK, gateway4 minutes agoProtected, verified
analytics-mcpMCP inspectionYesterdayNot governed
invoice-extractorRuntime GridProvider traffic seenSuspected bypass
Where it plugs in

Frameworks, SDKs and runtimes you already use.

python -m pip install agenticdome-python-sdk
agenticdome-demo --framework crewai --scenario refund_hijack

Requires Python 3.11+. Installation needs network access. The simulation then runs locally, with no account, API key or live payment.

Keep your native network clients, use SDK hooks for richer application context, and add Docker Agent’s native guard for local tools. See how the pathways work together →

Why action security matters now

Agents are arriving faster than controls.

AgentsTools and delegated agents create more execution paths.Connect the path. Verify the evidence.
MCPShared tools need a decision at each connected action boundary.Connect the path. Verify the evidence.
ActionsAgentic application risks include misused tools and delegated authority.Explore the public threat matrix and research.

Who it is for

Security leadersKnow which agents exist and prove which actions were allowed.
DevelopersAdd a check in a few lines, and test it locally first.
Platform teamsPlace the gateway on the right path and keep it running.
Risk and governanceEvidence for every decision, and review of what changed.

See an action blocked before it takes effect.

Try the local demo Book a deployment review