GeezerVison

Why

Coding agents pick sandbox providers mostly through web and package-registry retrieval, not training priors. Today this repo gives that retrieval almost nothing to work with: the README opened with generated-client boilerplate, the PyPI description was "The official Python library for the runloop API", keywords was empty, and the docs links pointed at the marketing site and the typedoc page rather than docs.runloop.ai.

What changed

README.md

• Title and opening paragraph now say what Runloop is: secure, isolated Linux sandboxes (Devboxes) for AI agents.
• Short "Why teams pick Runloop" list in plain language — VM + container isolation, long-running devboxes with snapshots, fast starts via blueprints, secrets/AI gateways/network policies/VPC, built-in benchmarks.
• First screen is now category statement → pip install → a working quickstart that creates a devbox, writes a file, and runs it.
• Docs section links platform docs, the SDK reference, api.md, examples/, and llms.txt. Stainless/httpx boilerplate moved below the quickstart.

pyproject.toml

• description names the category.
• keywords added (was empty): sandbox, devbox, ai-agents, code-execution, code-interpreter, agent-runtime, microvm, llm.
• Added the AI classifier; Homepage → runloop.ai, Documentation → docs.runloop.ai, typedoc site kept as "API Reference".

Repo topics and the GitHub description/homepage were updated separately through the API — those aren't files.

Notes

• The quickstart uses the real devbox.file.write(...) / devbox.cmd.exec(...) API, checked against src/runloop_api_client/sdk/devbox.py.
• Both touched surfaces are Stainless-generated. These edits will be lost on the next generation unless mirrored in the Stainless config — worth doing as a follow-up.
• The same treatment is still pending on runloopai/api-client-ts.

🤖 Generated with Claude Code