Spawn AI agents
from files.
Stack tools, skills, and memory into reproducible worlds. One file, runs anywhere. Open-source.
Models are getting smarter. The places they live in haven't caught up.
Read the manifesto→Two agents, two worlds.
Same project, different stacks. Every agent gets the sandbox its job needs, not the sandbox everyone else got.
dependencies: - "spwn:python" - "spwn:postgres" - "skill/tests"
dependencies: - "spwn:node" - "tool/ffmpeg" - "skill/encode"
One spwn up. Two containers. Zero overlap.
The source code of an environment.
Every block is a file. Agents declare what they need, worlds declare how they run. Stack them, override them, share them like Dockerfiles for minds.
“The next breakthrough isn't smarter models. It's richer worlds.”
The real power of AI isn't the model. It's the model plus everything around it. Oppenheimer in a chatbox can answer questions. Oppenheimer in a lab, surrounded by instruments, notebooks, colleagues, and years of memory, can change the world. The environment is the multiplier.
From your laptop down to the agent's core.
Every layer is a spwn contract. Host, architect, world, agent; declared in code, deployed by spwn up, torn down by spwn down.
Every block is a file.
Not a wrapper. Not a framework. Composable primitives a PR can review.
🧾Agents orchestration as code
Commit agents alongside your app, review behavior changes in PRs, ship the same mind to every machine.
🛠️Composable environment
Stack spwn:python, spwn:qmd, tool/ffmpeg, whatever your agent needs. Each one wakes up in a sandbox assembled for its job.
📦Reproducible
One spwn build, one portable artifact. Byte-identical on every machine, every time.
🪐Worlds
Bundle agents, workspaces, and knowledge. spwn up deploys them together, spwn down tears it all down.
🧠Persistent
Memory is a folder of markdown files. Readable, diffable, alive across restarts.
🧐Checked
spwn check surfaces bad refs, missing files, and lockfile drift before you spawn anything.