Show me

See the control loop before a bad action runs.

Pick an OpenClaw, Hermes, or custom-agent boundary, then inspect local block, live preflight, or sandbox gate responses, decision headers, and proof digest without exposing runtime secrets.

Inspectable demo surface

Live strict preflight

Calls the live runtime preflight endpoint and blocks before the function body.

public runtime call

ready

$ curl -i /api/public/demo/sessionpublic browser -> Cloudflare WorkerWorker holds credentials; browser receives redacted proof onlychoose a boundary above to populate headers

Runtimepending

Decisionpending

Proof digestsha256:pending

Worker timingpendingms

Headers to inspect

server-timing: run demo to populatex-imladri-worker-duration-ms: run demo to populatex-imladri-demo-runtime: run demo to populatex-imladri-demo-decision: run demo to populatex-imladri-demo-proof-digest: run demo to populatex-imladri-demo-glasshouse-status: run demo to populate

Five-step flow

From account setup to proof export.

Publish

Publish the policy

The customer chooses allowed, blocked, review-required, sandbox, and database boundaries for one OpenClaw, Hermes, or custom agent.

Wrap

Wrap dangerous capabilities

The agent keeps its normal brain and tools. Imladri sits around OpenClaw tools, Hermes plugins, custom SDK actions, database branches, cloud calls, payments, or compute jobs that can cause side effects.

Gate

Block or halt before the boundary

Known-bad local actions stop before network. High-risk actions can use live preflight, sandbox-before-spawn, or governed database execution.

Inspect

Inspect deployment state

The private pilot workspace keeps agent activity, blocked attempts, halt state, constitution revision, DB sandbox evidence, training runs, and proof export in one surface.

Export

Export SHA-256 proof

Security review gets a signed packet with the published policy, runtime decision, evidence rows, and proof integrity metadata.

Customer code shape

Wrap the tool body.

Imladri is not a replacement for the customer's agent. It is the runtime boundary around side-effectful OpenClaw tools, Hermes plugins, or custom SDK actions.

const guardedDeploy = agent.action(
  "deploy.preview",
  async (plan) => deployPreview(plan),
  {
    strictPreflight: true,
    intent: (plan) => ({ service: plan.service, version: plan.version }),
  },
);

await guardedDeploy(releasePlan);

Pilot evidence

The private workspace becomes the control room.

agentrelease-agentactionpayment.transferoutcomeBLOCKED before function bodyproofSHA-256 evidence packet

Security model

Honest boundaries are part of the product.

Preventable

Dangerous capabilities routed through SDK action wrappers, strict preflight, sandboxed actions, or governed database actions.

Observable and haltable

Capabilities not routed through the boundary can still become evidence and halt triggers, but they are not pre-execution preventable.

Live halt state

Operator halt propagates to SDK, strict-preflight, sandbox, and database enforcement boundaries before the next side-effectful action runs.

Execution lanes

Local blocks optimize known-bad actions. Live preflight checks current state. Sandbox and database lanes add process, branch, or SQL-scoped proof.

Stale cache behavior

Local cached constitution state can trail a new publish briefly. Force poll or live preflight is the stronger path for urgent changes.

Proof integrity

Customer evidence stays on the SHA-256/FIPS-compatible path. BLAKE3 remains experimental only.

Use cases

Start with one privileged workflow.

OpenClaw and Hermes agents

Connect the runtime bridge, publish policy, verify one allowed action, block one dangerous native tool or plugin body, and export proof.

Code and deploy agents

Wrap deploy.preview, release promotion, pull-request mutation, and infrastructure changes before they touch production.

Database agents

Branch Postgres sandboxes, route SQL through approved templates, isolate writes, and prove the source database stayed untouched.

Cloud and infra agents

Gate IAM, resource creation, billing-sensitive operations, and environment changes with halt and proof export attached.

Pilot path

Bring one OpenClaw, Hermes, or custom agent with one dangerous tool.

The useful pilot is narrow: publish policy, wrap the tool, run one allowed call or database branch, block one prohibited action, verify halt, and export proof.

Join the waitlist Contact us