BayLeaf is a situated counterplatform for Generative AI at UC Santa Cruz: an experimental service operated by Adam Smith (Dept. of Computational Media). It offers a working platform on different terms: local, bounded, revisable, and answerable to the contexts in which it is used.
BayLeaf has undergone a security review by UC Santa Cruz ITS: the campus Chief
Information Security Officer (CISO) and his team deemed it ready to launch,
safe for use by the UCSC campus community. This attests to its security posture
and is not an adoption of BayLeaf as an ITS-operated or ITS-supported
service; it remains faculty-operated. See SECURITY.md and
politics/SECURITY.md.
BayLeaf Chat — chat.bayleaf.dev
An Open WebUI deployment offering curated AI models to UCSC students, faculty, and staff. Features include:
- A Basic model backed by a rotating LLM and customized with a campus-aware system prompt
- Curated proprietary and open-weight models, all routed through OpenRouter ZDR
- Invite-code-gated groups for course-, department-, or role-specific models and toolkits
- Web Search and Web Page Content tools available to all users
- Per-turn rate limiting for fair, cost-efficient access
BayLeaf API — api.bayleaf.dev
An API that gives the campus community programmatic access to LLMs, web search and page fetching, and sandboxed code execution:
- Keyless inference and web-tool access from the campus network (169.233.x.x)
- API key access for off-campus use (self-issued via the service)
- Web search & fetch — search the web and extract clean page content from one or many URLs at a time, both backed by Tavily, available to all authenticated users
- Code sandbox — persistent Linux environments (backed by Daytona) for running code, uploading/downloading files, all authenticated with the same personal API key, including on campus
- Leaves system and developer instructions under the API caller's control
- Plaintext inference via OpenRouter, limited to models with a nonempty
hugging_face_idand a resolving Hugging Face repository. Missing, negative, or unavailable evidence fails closed with HTTP 403 - Sealed confidential inference via Tinfoil, whose current catalog lists only open-weight models. The requested model and body are encrypted from BayLeaf, though non-streaming usage metadata can report the executed model
Status — uptime dashboard
Public uptime dashboard for all BayLeaf services.
Questions, problems, or feature requests? Open an issue on this repo.
This is a small, faculty-operated project. Response times are best-effort, but every issue is read.
All active LLM inference uses zero-data-retention (ZDR) providers. Chat and the API's plaintext lane use OpenRouter; API Sealed uses Tinfoil, with the requested model and body encrypted from BayLeaf. No LLM provider stores prompts or completions. Vertex AI, AWS Bedrock, and NRP are disabled and carry no traffic. Non-AI features that require persistent storage, such as the Daytona-backed Code Sandbox, store user files by necessity.
api/— BayLeaf API Cloudflare Worker (api.bayleaf.dev)landing/— Static GitHub Pages site published at bayleaf.dev: one comprehensive landing page (privacy, support, and use cases live as GitHub-served markdown: PRIVACY.md, SUPPORT.md, USE-CASES.md)chat/— BayLeaf Chat DigitalOcean App (chat.bayleaf.dev): workspace models, custom tools, filters, and design docpolitics/— The case for universities owning their own AI infrastructure (manifesto)
This repo is publicly visible. It never contains API keys, credentials, or other sensitive configuration.
Nearly 100% of the code, documentation, and other project data in this repository was created using generative AI in agentic coding tools. This is an intentional choice: it demonstrates that sufficient technical capacity exists within the university to build and operate a service like this, without ceding control or responsibility to external parties. Critics, allies, and other humans seeking a direct human connection should contact Adam Smith directly.