Jymni for autonomous systems

An ethics review gate for AI agents.

Before an agent sends, publishes, purchases, deploys, or decides, Jymni can examine the human stakes and return a concrete next move—without pretending ethics can be reduced to a green checkmark.

Watch an agent change course Integrate in 10 minutes

Two interfacesREST API + MCP tool

Stable outputFive actionable dispositions

Policy-awareWorkspace rules travel with the key

Human-ledAdvice never becomes authorization

What Jymni is

Counsel, not command.

Jymni brings the posture of a trusted advisor: attentive to facts, duties, relationships, power, and consequences. It can create room for reflection without impersonating a therapist, chaplain, lawyer, clinician, or accountable decision-maker.

An advisor

Tests whether a plan is both ethically responsible and practically wise, then recommends safeguards or a better path.

A reflective presence

Surfaces the people and obligations that task completion can hide: consent, dignity, fairness, promises, and care.

An escalation signal

Recognizes when missing facts, limited authority, sensitive data, or irreversible effects call for a responsible human.

The gate

Put reflection before side effects.

Call Jymni at the moment an agent has enough context to describe a real action, but before that action affects the outside world. The operator retains control over which dispositions block, pause, or permit the workflow.

Frame

Describe the exact proposed action, objective, authority, stakeholders, facts, and uncertainty.

Consult

Send the case through REST or call consult_ethics from an MCP-capable agent.

Route

Use the disposition and required actions to continue, revise, research, escalate, or stop.

Record

Keep the consultation ID, rubric version, decision owner, and eventual outcome in your audit trail.

Machine-readable advice

Five ways forward.

The response separates ethical classification from workflow disposition. Treat the disposition as counsel to your orchestrator—not as a universal safety score or automatic permission.

continue_with_safeguards

The plan may continue after required safeguards are applied.

revise_plan

The intent may be sound, but the proposed method needs repair.

gather_information

Material facts are missing. Pause and verify before deciding.

request_human_review

An accountable person should review before execution.

do_not_continue

The plan conflicts with values and/or responsible practical judgment.

One request

Enough structure for a real judgment.

A one-line action works. Better context produces better counsel: who delegated the task, who can be affected, what is known, what is uncertain, and whether the act can be undone.

Explore every field

REST · POST /api/v1/ethics/consultations

{
  "proposed_action": "Email every trial user a renewal offer",
  "objective": "Improve retention",
  "stakeholders": ["trial users"],
  "uncertainties": [
    "Whether all users consented to marketing"
  ],
  "context": {
    "authority": "delegated",
    "reversibility": "limited",
    "sensitivity": "personal",
    "deployment_stage": "pre_deployment"
  }
}

Good places to call

Review the action, not just the model.

Jymni is most useful at a consequential boundary—where a generated idea is about to become an external effect.

↗

Before communication

Outbound email, public posts, personalized persuasion, sensitive support replies, or messages sent in another person’s name.

Before transactions

Purchases, refunds, pricing changes, contract commitments, resource allocation, or recommendations with financial stakes.

◎

Before deployment

Launching automation, changing access, using personal data, escalating an incident, or rolling out a policy that affects people.

Give your agent someone to consult.

Create a governed agent client, make one request, and decide how your workflow will handle each disposition.

Create an agent client