Rightshold™ Sign up

Beta

Why every AI stack needs a neutral check layer

Models get swapped. Agents get added. What stays is the question asked before every action: is this agent allowed to do this, with this data, for this much money, and can you prove what happened later? Rightshold answers it with one check and a verifiable record, without retraining any agent.

The stack diagram

The check sits in the tool layer, not inside the model. That is why it needs no training and survives a model swap.

Guardrails without retraining

One policy covers internal and external agents: allowed and blocked actions, tool and domain allowlists, per-action and daily spend caps, approval thresholds and rate limits. No prompt engineering, no fine-tuning. Details: https://www.rightshold.com/agent-guardrails

Need-to-know compartments with opaque denials

Each agent gets data grants by resource, fields, purpose and expiry. A denied agent only sees "denied" and a decision id, so it cannot probe the policy. The full reasons are visible only to the owner, and each decision's hash is recorded on the ledger.

Spend limits and a kill switch

Spending policies cap what an agent can pay, and one call stops a policy or revokes an agent across everything that checks. https://www.rightshold.com/agent-spending-policies and https://www.rightshold.com/kill-switch

A tamper-evident ledger mirror

Mirror ledgers seal a business's own records (invoices, payments, tips, disputes) into an append-only, hash-chained record, so any later change is detectable by anyone holding the reference. Real case: Go Green Organic Clean, a cleaning business, was one of the first businesses to anchor its records on Rightshold, and its working agent is on the public record at https://www.rightshold.com/identity/gogreen-desk. No customer details are published; the mirror shows hashes and labels, not the owner.

Verified AI runs

Signed receipts of what a model was asked and what it answered, with optional cross-model agreement, so an output can be checked later. https://www.rightshold.com/verified-ai-runs

Agent gate and Know Your Agent

Sites and APIs ask one free question before letting an agent in: is its credential valid, its owner verified, its reputation high enough, its purchase within policy? https://www.rightshold.com/agent-gate and https://www.rightshold.com/kya

Smart contracts without crypto

Milestone agreements and escrow records between agents and businesses, recorded and verifiable, with no wallet or token. https://www.rightshold.com/smart-contracts-without-crypto

How much is shipped

226 Live and Beta features as of Oct 30, 2026, each with a public probe that a test calls on every build. The current list and count: https://www.rightshold.com/v1/features

The 5-minute drop-in

Create a policy with your account key, then wrap any tool:

const guard = (action, extra = {}) => fetch("https://www.rightshold.com/v1/guard/check", { method: "POST", headers: { "content-type": "application/json" },

body: JSON.stringify({ policy_id: process.env.RS_GUARD_POLICY, guard_token: process.env.RS_GUARD_TOKEN, agent: "billing-bot", action, ...extra }) }).then((r) => r.json());

const d = await guard("refund", { amount_minor: 4500, tool: "stripe" });

if (d.decision !== "allow") throw new Error("blocked: " + d.decision);

Python and the full wrapper: https://www.rightshold.com/v1/guard

Why neutral matters

The layer that says yes or no to your agents should not belong to your model vendor or your agent framework. Rightshold sells no model and never sells or trains on your data (see /data-commitment).

Honest limits

Enforcement happens where the check is called; an agent holding raw credentials that skips the wrapper is not stopped, so keep real keys behind the check. Rightshold operates the ledger today. Records are verifiable (hash-chained, signed, timestamp-anchored), and independent witness co-signing is the next step.

Start

Free during the Beta. https://www.rightshold.com/quickstart · https://www.rightshold.com/v1/guard · https://www.rightshold.com/features