Ship agents your buyer can verify.

Every tool call signed, chained, and mapped to a control your buyer already audits.

Book a DemoSee how it works
LIVE · AGENT RECEIPTS

SOC 2 · HIPAA · EU AI Act · AIUC-1 · NIST AI RMF · SOX ITGC

Get your receipts started in an hour, prod by EOW.

Plan. Receipts. Verification.

Three steps: define what the agent is allowed to do, record what it actually did, and hand your buyer something they can verify themselves.

Define the least-privilege plan.

Set scopes, ceilings, and checkpoints once. The plan says which tools the agent can call, under what conditions, and when a human has to approve.

  • Which tools are allowed
  • What data or entities are in scope
  • Session or dollar ceilings
  • Which actions require human approval
  • Which actions must always be blocked

Plan file

What the agent is allowed to do

A small policy file defines the rules before the agent takes action. Same plan, every run.

read:patient_record → allowed in active encounter
submit:claim → allowed under session ceiling
issue:refund → checkpoint required

One plan, one receipt chain, one verification step — repeated across every agent you deploy.

Logs are the DIY way. Receipts are the straight-line way.

LOGS (LangSmith, Datadog, your stack)

Traces, logs, spans, dashboards — you can build evidence from them, but it means pulling the right slices, mapping them to controls, and explaining them every time someone asks “what happened here?” Your GRC tool can collect those logs, but it still relies on you to tell the story.

RECEIPTS (AgentMint)

Each agent action is captured once at the right level, already mapped to controls and signed when it happens. When someone asks “what happened here?”, you hand them one receipts file and a verify script they can run themselves.

Your internal stays the system of record. AgentMint just gives it agent-level evidence that doesn't require an engineer in the room.

Install to production in a week, with minimal work.

01

We hop on a call

About an hour. Your engineer runs pip install agentmint and agentmint init --write while we're talking. The CLI wraps the functions that call tools; no new services or infra.

02

We write the plan together

You say what “safe” and “scary” look like for this agent. I translate that into scopes, ceilings, and human checkpoints so your team doesn't inherit a new governance system.

03

We test on real receipts

We run in shadow mode on real traffic. When a receipt looks wrong or surprising, you flag it; I update the plan and push the change. You're not maintaining anything — just saying “good” or “bad.”

04

Green light → production

Ship to prod. We will be there for your users. By EOW you have a receipts ZIP from real traffic plus a verify script your buyer can run offline.

05

I show up for your buyer

When their CISO has questions, I'm on the call so you don't have to improvise. We talk in their language — AIUC-1, EU AI Act, FedRAMP — using receipts instead of screenshots.

Worst case, you uninstall after a week and your agent goes back to how it ran before. You still keep a week of independently verifiable receipts as free insurance.

Get started.

FREE

The library. MIT licensed.

You self-host. You wrap your tool calls. Forever free.

PILOT

I instrument it with you.

One agent, one environment, one week in shadow mode. We hop on a call, wire it in, and write the plan together. By EOW you're in production and looking at receipts from real traffic.

Talk to me →

Custom tiers, volume pricing, and channel deals available.

Common questions.

Q

Where does this run?

Your machine, your process, your container. AgentMint sits inside your agent; nothing leaves unless you push it. The private key you use to sign receipts never touches any service — it's yours, on your infra. The corresponding public key is what you share with a buyer so they can verify.

Q

What if someone edits the receipt file after the fact?

The signature breaks. Each receipt is signed at creation time using Ed25519, and the chain links each receipt to the previous hash. Alter one byte, and verification fails instantly.

Q

We already use LangSmith / Datadog / Observe. Why add this?

Those are great for you; this is for your buyer. LangSmith and Datadog give you debugging traces and operational dashboards. AgentMint produces a single evidence artifact your buyer can verify themselves without needing access to your dashboards or trusting your ops narrative.

Q

Can I run this without the cloud?

Yes, and that's the point. The free library works entirely offline; you sign locally, you store locally, you verify locally. There's no phone-home, no telemetry, no SaaS dashboard. If you want a managed dashboard or long-term storage later, that option exists, but it's never required.

Still have questions? Let's talk.

The agentic internet needs an independent notary.

As AI agents handle money, healthcare, and legal decisions, every company using them will need verifiable evidence that those agents behaved correctly. Today that evidence is built by hand — logs, screenshots, and internal narratives. AgentMint replaces the manual work with a signed, chained, control-mapped receipt that any third party can verify without trusting you.

Open Source

SOC 2 · HIPAA · EU AI Act · AIUC-1 · NIST AI RMF · SOX ITGC · OWASP AST10 Listed (Ken Huang)

Full control mapping →