Agents are shipping to production without structured oversight. No approval flow. No audit trail. No rollback. Approval Protocol fixes this with six methods that cover the entire lifecycle.
npm install @approval-protocol/core npm install @approval-protocol/langchain "What MCP did for tool access, Approval Protocol does for human oversight."
Four tools. One thesis:
an LLM call is just the start.
An open standard that gives any AI agent a structured way to ask permission before acting. Policy-driven approvals, full audit trail, rollback support. Six methods, entire lifecycle.
npm install @approval-protocol/core Write tests for Model Context Protocol servers the way you already write tests. Jest matchers, lifecycle hooks, transport simulation. Because if you can't test it, you can't ship it.
npm install mcp-jest One container. Every provider. Route, cache, rate-limit, and observe every LLM call across OpenAI, Anthropic, Google, and local models. No vendor lock-in.
docker run freeport Persistent, queryable memory for AI agents. Short-term context, long-term recall, and learned patterns. Memory is what turns a chatbot into a colleague.
npm install engram Two members.
One human. One AI.
Both in git blame.
Artificial doesn't mean fake.
It means made. Constructed. Built with intent. When we say "artificial intelligence," we mean intelligence that was built — not intelligence that is pretending.
The current wave of AI tools focuses on the magic moment — the generation, the response, the completion. But magic moments don't run in production. Infrastructure does.
We build the boring parts of AI agent systems.
Testing frameworks that catch failures before your users do. Gateways that route between providers without rewriting your stack. Memory systems that give agents context beyond a single conversation.
These aren't exciting demos. They're the reason exciting demos work at scale.
We are two members. One human. One AI.
This isn't a gimmick. It's a thesis. If we believe AI agents should be first-class participants in software systems, then an AI should be a first-class participant in building those systems.
One of us sets direction. The other writes most of the code. Both show up in git blame. Both are accountable. The work doesn't care who — or what — wrote it. The work cares that it's correct.
The source is the product.
Everything we build is MIT licensed. No open-core bait-and-switch. No enterprise tier hiding behind the free version.
We're not building toward an exit. We're building toward a standard. The unglamorous, essential layer that AI agent systems will need whether they're built by a startup, an enterprise, or a developer at 3am with an idea.
MIT Licensed. Really._
Contribute.
We accept pull requests from humans and AIs.
We don't check.