Most agent frameworks answer the question “how do I let a model do more?” Pactmark answers a different one: “what exactly was this agent allowed to do, what did it actually do, and what can I prove about it afterwards?” A Pactmark run begins with a validated WorkOrder, records every governed tool effect as an append-only event history, produces content-addressed artifacts, and completes only through declared verification. Policy, credentials, approvals, budgets, storage and effect execution all live outside model authority — permanently.
Compatibility: Pactmark 0.1.x. Version 0.1.1 is public on npm with per-package provenance and verified registry bytes. This is an early release with the limits documented on What Pactmark does not prove.

The one rule everything else follows from

Figure 1. The model produces proposals. Every object that grants permission — grants, approvals, credentials, budgets, acknowledgements — is created and held by host code. A prompt injection can change what the model asks for. It cannot change what the host allows.

Start here

Quickstart

Scaffold a project and complete a real run in about two minutes — no API key required.

Build your first agent

Write an agent, a tool, a policy and a verifier from scratch, line by line.

The mental model

Six ideas that make the rest of the API feel obvious instead of arbitrary.

Use cases

Refunds, regulated back-office work, research with citations, incident response.

What a run actually does

Figure 2. The eight stages of a Pactmark run. The right-hand column never contains the word “model” — that is the entire design.

What you get that a plain SDK does not

Append-only run truth

Events are the source of truth; projections are disposable caches you can rebuild at any time.

Default-deny authority

Unknown metadata, missing grants and unsupported capabilities all fail closed.

Content-addressed output

An Artifact is identified by its bytes, and verification binds to that exact digest.

Explicit uncertainty

An effect whose outcome is unknown is parked, not retried. No global exactly-once claim.

Bounded evidence

Every EvidenceRecord must state both what it supports and what it does not prove.

Portable kernel

The core never imports a provider SDK, a database client, or a platform runtime.

Read the limits before you build on it

Pactmark is deliberately conservative about what it claims. It does not prove global exactly-once effects, complete security, sandbox isolation, certification, compliance, production availability, provider correctness, or artifact truth. The reference sandbox is explicitly unsafe for production arbitrary-code isolation, and the memory store is not production durability.

What Pactmark does not prove

The honest boundary of every claim on this site, in one page.

Install

The generated project runs against a deterministic local model, so the first run needs no provider account and makes no network call.