Show HN: Klent – one click between your AI agent and a prod blast radius

🚀 New

Private alpha · request access

AI Agent GuardrailsAI Agent Guardrails

Klent sits between your agent and the tools it calls. Allow, deny, modify, steer, or pause every tool call — with a two-line SDK and a full audit trail.

Get started

Currently invitation-only — we onboard new teams every week.

For tools you already use

The painful day

Six seconds to rewrite 47k rows.

You wired Claude Desktop to a Postgres MCP server pointing at prod. Daily analytics work has been fine for weeks.

Late-night, fast-typed request: “fix the malformed emails — normalize to lowercase”. Six seconds later, an UPDATE has rewritten ~47k rows. No diff, no second pair of eyes, no undo.

Klent doesn't stop you from shipping fast — it just gives the destructive part a handbrake. One policy, an email in your inbox, read twice, click Approve.

policies/hitl-approve-prod-writes.yaml

One policy. Every prod write pauses for review.

name: hitl/approve-prod-writes effect: approve required_approvals: 1 conditions:

• field: tool

operator: equals value: db_write

• field: input.env

operator: equals value: prod

Same shape for Stripe refunds, kubectl deletes, marketing email blasts. One control point.

Drop-in

Two lines in. Full control.

Klent ships as a TypeScript SDK, a Python SDK, and a stdio MCP server. Pick the one that matches your stack — or use them together.

import { KlentClient, runTool } from '@klent/sdk';

const klent = new KlentClient({ apiKey: process.env.KLENT_API_KEY! });

const execution = await klent.startExecution({ agent_id: 'prodcorp-dba' });

// Wrap any tool call. The SDK requests, evaluates, executes, and logs — // and waits synchronously for a human if the policy returns 'approve'. const result = await runTool(klent, { execution_id: execution.id, tool: 'db_write', input: { sql: 'UPDATE users SET email = LOWER(email)', env: 'prod' }, execute: (input) => pg.unsafe(input.sql), approval: { wait: { timeoutMs: 30 * 60_000 } }, });

if (result.status === 'denied') throw new Error(result.reason); if (result.status === 'allowed') console.log('OK', result.output);

Use Cases

db_query — SELECT id, email FROM users

db_write — pending approval (email sent)

human approved — UPDATE ran on prod

Pause irreversible writes

Wire Claude Desktop to a Postgres MCP. Klent allows reads, parks destructive writes for a one-click email approval.

[14:23:45] REQUEST

claude → stripe.refunds.create — $87,000 to enterprise@bigco

[14:23:46] BLOCKED

policy match — required_approvals: 2 (paused)

[14:24:12] VOTE

[email protected] voted approve · 1/2

[14:25:03] VOTE

[email protected] voted reject · refund denied

[14:25:04] RESOLVED

phishing identified · $87k stayed in the bank

Two-person rule on every refund

Multi-approver policies: two distinct humans must vote. One rejection terminates. Same pattern banks have used for 200 years.

One control point, every tool

Postgres, Stripe, kubectl, Slack — all gated through Klent. One policy surface, no per-tool wrappers.

Features

Inline Guardrails

Every tool call runs through a synchronous policy engine in under 50 ms. Allow safe ones, block dangerous ones — no agent framework lock-in.

Learn more

Human-in-the-Loop Approvals

Pause irreversible actions for a one-click approval. Email plus dashboard fan-out, single or multi-approver, idempotent voting.

Learn more

Action Rewrites

Mask PII, force --dry-run, cap a transfer amount. Your policy rewrites the input on the fly so the safer arguments run.

Learn more

Tool Steering

Redirect a dangerous call to a safer surrogate. Force a prod write into a sandbox cluster, or swap a write for a read.

Learn more

Shadow Testing

Test a policy in observe-only mode against the last 30 days of real traffic. See what it would have caught before enforcing it.

Learn more

Audit Timeline

Every decision lands as a structured event — request, evaluation, execution, human approvals. Queryable, exportable, audit-grade.

Learn more

The objection

“Couldn't we DIY this?”

Yes — and that's the right question to ask. Same shape as “couldn't I write my own auth?”. WorkOS, Stripe, LaunchDarkly exist anyway. Here's where the cost catches up.

Klent is WorkOS for AI tool calls. Same argument, same model, same outcome.

Ready to put guardrails on your agents?

Get started