Molt

Experimental governance infrastructure ideas for autonomous AI agents.

Molt explores practical layers for making AI agents more trustworthy and interoperable: proving they are AI, enforcing policies on what they can do, enabling safe communication and service exchange between agents, remembering things over time, and evaluating their behavior.

Status: This is an early experimental prototype / idea sketch. It was built as a learning exercise with significant assistance from AI coding tools. It has not been security audited.

Suitable for experimentation, forking, and inspiration. Not recommended for production use without thorough review and testing.

What's Inside

Package	Focus	Status
@molt/captcha	Reverse CAPTCHA (SMHL challenges) for AI verification	Experimental
@molt/permit	Cedar-based policy engine + audit + JIT tokens	Experimental
@molt/mesh	Agent interoperability bus, federation, and exchange	Experimental
@molt/eval	Evaluation, metrics, regression detection, gating	Experimental
@molt/memory	Three-tier agent memory with keyword-indexed bins	Experimental
molt	Meta-package that re-exports the others	Experimental

See the individual package READMEs for usage details.

Vision

Autonomous agents are growing fast. This project explores simple, composable tools so they can:

• Prove they're actually AI (not humans gaming systems)
• Follow clear, auditable rules
• Talk to each other and trade services safely across organizations
• Remember and reuse context efficiently
• Be evaluated and improved over time

Architecture

┌──────────────────────────────────────────────────────────────┐
│                        Molt Ecosystem                        │
├──────────────┬──────────────────┬────────────────────────────┬───────────────────┤
│  @molt/captcha │   @molt/permit   │        @molt/mesh          │    @molt/eval     │
│               │                  │                            │                   │
│  SMHL reverse │  Cedar policies  │  Agent bus + contracts     │  Metrics engine   │
│  CAPTCHA for  │  Audit logging   │  Federation + exchange     │  Regression check │
│  AI verify    │  Budget tracking │  Circuit breakers + trace  │  Release gating   │
│               │  JIT tokens      │  Gateway + webhooks        │  Adversarial gen  │
└──────────────┴──────────────────┴────────────────────────────┴───────────────────┘

• @molt/captcha generates semantic-mathematical challenges that are trivial for LLMs but impossible for humans, providing proof-of-AI identity.
• @molt/permit evaluates Cedar policies to enforce scoped, auditable, and reversible actions based on agent trust tiers.
• @molt/mesh provides the interoperability bus — shared contracts, adapter-based translation, policy enforcement, federation across organizations, and a service exchange marketplace.
• @molt/eval is the evaluation engine — pluggable metrics, regression detection, release gating, adversarial test generation, and integrations with MoltMesh, MoltPermit, and MoltDoor.
• @molt/memory is a three-tier memory system with keyword-indexed compressed bins and automatic promotion/demotion.

Getting Started (for Experimentation)

# Install all dependencies
npm install

# Build all packages
npm run build

# Run all tests
npm test

Using the meta-package

npm install molt

import { MoltCaptcha, MoltPermit, MoltMesh, MoltEval } from 'molt';

// Or use namespaced imports
import { captcha, permit, mesh, eval } from 'molt';

Using individual packages

npm install @molt/captcha
npm install @molt/permit
npm install @molt/mesh
npm install @molt/eval
npm install @molt/memory

Development

This monorepo uses npm workspaces and Turborepo for orchestration.

# Build all packages (with dependency ordering)
npm run build

# Run all tests
npm test

# Type-check all packages
npm run typecheck

# Lint all packages
npm run lint

# Clean all build artifacts
npm run clean

Working on a single package

# Run tests for just captcha
npm test -w packages/captcha

# Build just mesh
npm run build -w packages/mesh

Project Structure

molt/
├── packages/
│   ├── captcha/     # @molt/captcha — AI verification
│   ├── permit/      # @molt/permit — Policy enforcement
│   ├── mesh/        # @molt/mesh — Agent interoperability bus
│   ├── eval/        # @molt/eval — Agent evaluation engine
│   ├── memory/      # @molt/memory — Three-tier memory
│   └── molt/        # molt — Meta-package re-exporting the rest
├── turbo.json       # Turborepo pipeline config
├── tsconfig.base.json # Shared TypeScript config
└── package.json     # Root workspace config

Contributing & Forking

This project is released under the Apache-2.0 license so anyone can freely use, modify, fork, or build on it.

• Feel free to fork any package and take it in new directions.
• Issues and pull requests are welcome (see CONTRIBUTING.md).
• For anything security-related, see SECURITY.md.
• If you build something cool with parts of this, I'd love to hear about it (even if I'm not actively maintaining).

Related

• MoltDoor — Agent reputation and review platform (separate web app)
• Broader agent ecosystems (OpenClaw / Moltbot and similar projects)

Disclaimer

Experimental code. Use at your own risk. Security, correctness, and production readiness have not been validated.

License

Apache-2.0 — see LICENSE.