AI that follows your teaching, and shows its work.
VIABLE is shared AI infrastructure for learning and research: grounded in your materials, checked against your guidance, on any model, and on the record.
Teachers and librariansTurn your materials into an assistant that cites them.
Education developersOne OpenAI-compatible API for agents and any model.
ResearchersStudy real use, with every turn, step, and cost on record.
Course sites, library guides, PDFs, slides, and notes.
Knowledge
Start with what you already trust.
Add course sites, library guides, PDFs, slides, spreadsheets, or pasted notes. Websites are checked again on the schedule you choose, and search reads whole sections while leaving out the menus a site repeats on every page.
Import a whole website, specific pages, or every page linked from one, with LibGuides sub-pages included.
A concept map shows the themes your sources cover, and where coverage is thin.
Uploaded files stay private unless you let readers open what an answer cites.
guides.example.eduWhole website, with sub-pages. Checks weekly.214 pages
catalog.example.edu/hist-210Specific pages3 pages
Files
research-methods-syllabus.pdf12 pages
week-3-primary-sources.pptx18 slides
Office hours notesPasted text
Map
Finding sourcesEvaluating evidencePrimary sourcesCitation stylesData and statistics
Thin coverage Patents, oral histories
Agent designer
Describe the agent. Get a working draft.
Write what the agent should do and choose the knowledge it answers from. The designer drafts the workflow, checks it, and opens it in the editor. When it answers from your sources, every factual sentence cites one.
It reads any page you mention while it drafts.
Or start from a template: a course help desk, a learning coach with memory, a tutor with guardrails, a VETTING tutor, and more.
Answer research questions from our library guides and course files.Cite every source, and suggest asking a librarian when the guides don’t cover it.
Answer from these knowledge bases
Library research help
Build agentDrafted, checked, and opened in the editor
InputinputPatron question
AgentresearchSearch and read knowledge
ReplyreplyWith sources
Agents
Refine it together, on any model.
Teammates and classmates edit the same canvas live and see who is changing what. Steps can search knowledge, remember each learner, choose between paths, run in parallel, read a web page, or call another agent, and each step picks its own model.
The assistant reads recent runs and proposes changes you apply or discard.
Memory keeps what each learner shared across conversations, and long conversations can be summarized as they near a model’s limit.
Open-weight models hosted by VIABLE sit beside frontier models in every step’s model list.
Three recent runs answered only from the library guides. Search the course files too?
Apply to draftDiscard
Applied. Review it before you publish.
VETTING and guardrails
Check every answer before a learner sees it.
Guardrail steps catch personal data, jailbreaks, harmful content, and off-topic questions. VETTING, short for Verification and Evaluation Tool for Targeting Invalid Narrative Generation, has a second model review each draft against your guidance or sources and send it back until one passes.
In a peer-reviewed study with 151 middle school students, VETTING stopped an estimated 91% of inappropriate drafts before students saw them, for about 20% more tokens.
Answers list their sources, down to the page of a PDF.
Publishing first shows what changed since the live version, step by step.
/v1/chat/vetting and /v1/vetting/full bring the same checks to your own requests.
Where can I find newspapers from 1918 about the flu?
Start with the library’s historical newspaper databases 1. The history course lists free digitized archives by state 2.
1Primary sources guide Newspapers
2research-methods-syllabus.pdf p. 4
InputPatron question
GuardrailNo personal data
Agent3 tool calls: search, read, read
VETTINGAttempt 1: a claim isn’t in the sources
VETTINGAttempt 2 passed
Reply2 sources
Chat pages
Live in minutes, where learners already are.
Turn on a chat page and share its link or QR code, or paste one line into LibGuides, LibAnswers, or any website. Canvas takes an inline frame. Visitors need no account or key, and every reply shows its sources.
Set the look, greeting, suggested questions, daily and per-visitor limits, and the sites allowed to embed it.
Replies can include tables, math, code, and diagrams.
Your own apps call the same agent with /runs, a Chat Completions request, or background jobs.
Every run is recorded step by step, with each model call, its tokens, its cost, and who paid. Tag calls by study, feature, and session, then filter, replay, and export them as CSV or JSONL.
Images and audio are archived byte for byte, opened only through signed-in links.
Authorization headers, API keys, tokens, and passwords are always redacted.
Each key has its own scopes and rate, token, cost, and model limits.
Run at 14:06 Guardrail, Agent with 3 tool calls, VETTING in 2 attempts, Reply
Payloads kept for replayCredentials redacted
Example rowsExport CSVExport JSONL
API and CLI
Or build it all in code.
One OpenAI-compatible API reaches frontier models and the open-weight models VIABLE hosts, with one key and one base URL. Bring your own provider keys when you want: requests keep the same VIABLE key, and a key that runs out hands the request to the next one in its pool.
Provider keys work with OpenAI, Google Gemini, Anthropic, OpenRouter, Azure OpenAI, Groq, self-hosted endpoints such as vLLM or Ollama, and more.
Paste one message into Claude Code, Codex, or another coding agent, and it sets up the VIABLE CLI and its Agent Skill.
Message for your coding agent
Read https://www.viablelab.org/downloads/viable-cli.md and set up VIABLE CLI and its Agent Skill for your environment. Help me sign in securely.
Read https://www.viablelab.org/downloads/viable-cli.md and set up VIABLE CLI and its Agent Skill for your environment. Help me sign in securely.
# Sign in in the browser. No password in the chat.$ viable auth login --device
# Create a scoped key straight into a file the agent never reads$ viable developer keys create --file key.json --output ./viable-api-key.txt --confirm create-key
# Preview an agent’s model calls before paying for a run$ viable developer agents test-run <agent-id> --message "Hi" --dry-run
intro-researchKnowledgeLibrary research help
Library research helpHybrid search
Websites
guides.example.eduWhole website, with sub-pages. Checks weekly.214 pages
catalog.example.edu/hist-210Specific pages3 pages
Files
research-methods-syllabus.pdf12 pages
week-3-primary-sources.pptx18 slides
Office hours notesPasted text
Map
Finding sourcesEvaluating evidencePrimary sourcesCitation stylesData and statistics
Thin coverage Patents, oral histories
Made by people who study learning.
VIABLE Lab, a learning-science research lab at the University of Florida, built the platform for its own studies and products, and runs it as shared infrastructure for other teams building learning tools.
This page is set in Atkinson Hyperlegible Next, a typeface from the Braille Institute designed so every letter is easy to tell apart.
Questions, answered.
The API reference covers every endpoint, and the API guide is written for agents and people alike.
What is the VIABLE Developer Platform?
Shared AI infrastructure from VIABLE Lab for learning products and research. It brings frontier and open-weight models into one OpenAI-compatible API, with projects, scoped keys, collaborative Agents, VETTING, and request logs, so a team can take a model experiment to a traceable workflow, research tool, or integrated application without rebuilding its stack.
What is the VIABLE API?
An OpenAI-compatible AI API for learning products and research. One key and one base URL (https://api.viablelab.org) reach frontier models and open-weight models hosted by VIABLE, plus capabilities built for education: VETTING verification, collaborative Agent workflows published as endpoints, and research-grade request logs.
Can students or library patrons use an agent without an app?
Yes. Turn on an agent’s chat page in its Connect tab and share the link, or paste one line into LibGuides, LibAnswers or any website to add a chat button; Canvas pages take an inline frame. Visitors need no account or key: the page runs on a project key you choose, within a daily limit you set, and every answer shows its sources.
What can an agent use besides a model?
Knowledge bases (websites crawled on a schedule, uploaded files and pasted text) that it searches and cites, memory of each user across conversations, web pages it reads, and other published agents. Each run records every step and model call, so you can see why it answered as it did.
How is it different from an API gateway or model router?
Gateways stop at model access. The VIABLE API adds VETTING, which verifies replies against pedagogical guidance or answer keys before they are returned; Agents, which teams build together and publish as one endpoint; and request logs designed for analysis, with replayable payloads and exports.
Does it work with the OpenAI SDK?
Yes, and it does not require it. Set the SDK base URL to https://api.viablelab.org/v1 and use a VIABLE API key, or send the same JSON with any HTTP client. Chat completions, embeddings, image generation, speech, and transcription keep OpenAI request and response shapes for every model in the catalog.
What is VETTING?
VETTING (Verification and Evaluation Tool for Targeting Invalid Narrative Generation) is a verification layer over any chat model. A model drafts a reply, a second model checks it against your hidden guidance or an answer key, and failed drafts are revised with the verifier feedback until one passes or the attempt limit is reached. Each response reports how many attempts ran and whether the final reply passed. In a peer-reviewed study with 151 middle school students, it stopped an estimated 91% of inappropriate drafts before students saw them.
Yes. Invite people to a shared project and everyone edits the same Agent live, with cursors, presence, automatic saves, and restorable versions. Owners manage members and keys; editors build and test; viewers follow along. It works well for teaching agent design as a group project.
Can a coding agent set it up for me?
Yes. Give Claude Code, Codex, or another agent with a terminal one message pointing at the VIABLE CLI setup guide. It installs the CLI and its Agent Skill, helps you sign in in your browser, and then works from the maintained API guide. Keys it creates are written to a file, not shown in chat.
What happens to request data?
Request metadata, usage, cost estimates, and request and response text are kept by default so requests can be replayed, filtered, and exported from Request Logs. Authorization headers, API keys, tokens, and passwords are always redacted. The VIABLE Lab privacy policy describes how logged data is handled.
Who operates the VIABLE API?
VIABLE Lab, a learning-science research lab at the University of Florida, builds and operates it. The public contract is published as OpenAPI, and the developer platform, CLI, and model catalog are documented on developers.viablelab.org.
