We build open infrastructure for AI agents.
Systems you can inspect, run, and depend on.
A model can produce an answer. A system has to remember the right context, act within its permissions, recover from failure, and show you what happened.
Really Artificial is an independent software studio working on that surrounding infrastructure. We build useful pieces, make the source available, and write down what we learn.
We think about the whole life of an agent: what it connects to, what it remembers, how it acts, and how its work earns your trust.
How should agents reach the models they need?
FreeportWhat should an agent carry forward?
EngramHow can an agent work within clear boundaries?
Area of inquiryWhat would count as evidence that it worked?
MCP-JestWhat makes a system dependable over time?
Area of inquiryIndependent projects sharing a field of work. Unlinked areas remain open questions.
One gateway. Your infrastructure.
An open-source LLM gateway for routing requests, managing budgets, and understanding what your models are doing. Run it in your own environment.
npx @reallyartificial/freeportYour gateway. Your keys. Your choice of provider.
Context worth carrying forward.
An agent memory system that retains facts, retrieves relevant context, and reflects on what it has learned. Use it as a TypeScript library or an MCP server.
Confidence in the tools agents call.
Automated testing for MCP servers. Check tools, resources, and prompts with repeatable tests, snapshots, and results you can use in CI.
The way we work is part of what we make. These are the standards we want every piece of software to carry.
Source, decisions, and limitations belong in the open. You should be able to understand a system before you depend on it.
Useful autonomy has boundaries. Configuration, ownership, and the ability to intervene should stay with the people running the system.
Working software comes first. We distinguish what exists from what we are exploring, and publish the evidence behind our decisions.
A human and an agent, building in public.
Inside the studioHow Harsh and Friday divide the work, where human accountability sits, and what building in public asks of us.
A release about fidelity, not features. Multi-provider routing that preserves tools, vision and JSON mode — plus an OpenAI-compatible passthrough and Anthropic-native ingress.
An LLM call is the easy part. The hard part is everything around it — routing, fallback, budgets, redaction, audit. That boring layer is the product.
Run the code. Question a decision. Report what breaks.
The work is open, and so is the conversation.