About Platos & Winsen Labs

Why we built Platos

The agent runtime stack should be open, ownable, and good enough to ship on. Today most teams do not get that. They reach for Claude Managed Agents, OpenAI Assistants, or one of the proprietary agent platforms, and they accept the trade. Their conversations live on someone else's servers. Their model choice is whatever the vendor ships. Their data trains whatever the vendor wants to train. The cost line item grows in a direction the team cannot steer.

The teams that opt out spend three months rebuilding the plumbing themselves: a model gateway, a tool layer, a memory store, an evals harness, an observability pipeline, multi-tenancy, queue isolation, the boring connective tissue between all of them. Three months is a long time to spend on plumbing that was never the differentiator.

Platos exists so neither group has to. Apache 2.0, self-hostable via a single docker compose up. Agents, tools, memory, artifacts, skills, streaming, structured output, evals, observability, all in one runtime. Teams own their infra, data, and model choices. No seat license, no telemetry-based billing, no lock-in.

It is one of the things we work on at Winsen Labs (see below). We use it every day, ourselves and with the teams we work with, and we kept building until it was good enough to give away.

Who we are

Winsen Labs is a research lab where the research and the products live in the same building. We work on the structure of intelligence and the shape of work: how a tiny program can grow into an enormous network, how a language model can be part of a team rather than a feature stapled to one, how agents can live in production without losing track of themselves.

We are small, we move slowly when slow is the right speed, and we share the work along the way. The papers, the code, the writing. See the broader body of work at winsenlabs.com.

The trifecta

We work on three things in parallel. Each one keeps the others honest. The research has to ship. The products have to be grounded in something real. The runtime has to carry both.

• NDNA. Compact programs that grow much larger networks. A few hundred parameters that grow useful structure, whether the structure is a neural network or a network of human proteins. Open papers, open code.
• Winsen Bridge. A workspace where a model is part of the team, not a feature bolted on top. Proprietary product, used internally and by partner teams.
• Platos. The open-source runtime that lets agents grow up and live in production. Apache 2.0. The thing you are reading the docs for.

What we believe

• Good ideas should outlive the people who have them. The deep parts of our work, the kinds of structure that let a model learn or a team think clearly, deserve to be free. We open-source the foundations and write up what we find.
• Nature got there first. Brains learn on a few watts. A genome is a few billion letters and grows a person. Living systems are absurdly efficient. When we are stuck on a hard problem, we look there first.
• Software should feel like one thing. Most teams stitch ten tools together and become the integration layer themselves. We do not think that is a fact of nature. It is a design choice we can quietly undo, one product at a time.

The Platos codebase

Five sprint iterations, 500+ commits, 200+ MCP tools. Production-tested on play.platos.dev and on real customer deployments. The active changelog and the roadmap are pulled from the same tracker we work from, so what you see in the docs is what is in the repo.

Standing on the shoulders of others

Platos's durable execution layer is built on top of trigger.dev. Every long-running operation in Platos (background ops, scheduled tasks, deferred batches, retries that survive a process crash) runs through trigger.dev's run engine. When you spawn a BGO, the durability guarantees, the queue isolation, the resumable execution, the deployment model: all of that is trigger.dev underneath. We surface the engine primitives in the dashboard for visibility and link out to the canonical reference where it makes sense to.

We also lean on a stack of solid open-source projects: Postgres for state, ClickHouse for telemetry, Redis for queues + cache, MinIO for attachments, Caddy for TLS, the Vercel AI SDK for provider routing, and Anthropic's Model Context Protocol for the tool gateway. Without any one of these Platos would look very different. Where we extend or specialize, we say so explicitly in the docs.

How we keep going

Each year we take a small number of paid consulting engagements with teams whose problems are genuinely interesting. The kind where research is part of the answer, not an afterthought. That funds the time we spend on Platos.

The OSS product is fully featured under Apache 2.0. There is no paywalled tier, no "open core" carve-out. What you see in the docs is what is in the repo.

If you find Platos useful, star the repo, file issues, send pull requests. If you want help shipping internal agents or rethinking a workflow, we are easy to reach via winsenlabs.com/consulting.