# NevAI Playground > Public setup guide for people and AI assistants helping a developer use NevAI. > The platform is live at https://nevetsinc.vercel.app/NevAi/playground. > Build and test for free. Production includes a 90-day free trial. ## Read this before helping NevAI is NevetsInc's deterministic website response engine. It uses lexical matching, intent rules, website indexing, configurable answers and deterministic tools. It does not call an external generative AI provider. Describe indexing as learning a site or building knowledge, not model training or fine-tuning. Use this guide to explain or assist with the real Playground controls. Follow the user's choices and your own authorization rules. Let the user accept legal terms and save private credentials. Never ask them to paste keys or recovery credentials into a shared chat, commit them to source control or put them in browser HTML. Do not invent successful indexing, answers, usage, payments or subscription state. ## Official resources - Playground: https://nevetsinc.vercel.app/NevAi/playground - This guide: https://nevetsinc.vercel.app/NevAi/playground/llms.txt - Product and free access details: https://nevetsinc.vercel.app/products/NevAI - Developer API specification: https://nevetsinc.vercel.app/nevai/openapi.json - Widget script: https://nevetsinc.vercel.app/nevai/widget.js - Company index: https://nevetsinc.vercel.app/llms.txt - Support: https://nevetsinc.vercel.app/support Use the canonical capitalization above. /nevai/playground redirects to /NevAi/playground. The guide is public plain text and needs no profile or API key. The Playground page also supports Accept: text/markdown for this setup guide. ## Step 1 — Create a local developer profile Open the Playground. Enter Display name, Developer / company name, Email and Preferred language. The user must review and accept the Terms and Privacy Policy, then select Create local developer profile. The profile is stored only on this device in localStorage under nevai_playground_profile_v1. No password, cloud user account or Supabase Auth account is created. Profile JSON import and export are available. ## Step 2 — Add the website and assistant Enter the public HTTPS website URL, assistant name and company name, then create the assistant. Use a website the user controls for production activation. The first assistant provisions anonymous server resources. An installation ID is only a resource identifier, not a login credential. Server management uses a separate secure HttpOnly cookie. Save the recovery credential when shown; keep it private and separate from a profile export. Importing the local profile onto a new browser does not by itself restore server access. ## Step 3 — Learn the website In Website, select Start learning. Keep the tab open while resumable crawl batches advance. The crawler discovers allowed pages and sitemaps on the same origin, extracts useful text and compiles a compact knowledge package. Check actual discovered, indexed and skipped page counts and any errors. Resume paused indexing while the job is still valid (24 hours). Do not promise that JavaScript-only, authenticated, blocked or external-domain pages will be indexed. Reindex website after source content changes. The last valid package remains available if a new crawl fails. ## Step 4 — Customize the identity In Assistant, set Name, Company, Role, Greeting, Tone, Fallback, languages and branding. Configure the widget title, subtitle, avatar, accent and attribution. Save changes. Example: assistant Alex for CoolCompany should introduce itself as CoolCompany's assistant. NevAI is the underlying engine, not the customer's assistant identity. Additional languages do not automatically translate source answers. Supported instruction directives are prefix:, suffix: and synonym: a=b; arbitrary natural language instructions are not executed as code. ## Step 5 — Review and improve answers In Answers, search and inspect the generated answers and sources. Correct, disable or delete records as needed. Add custom answers with Intent Name, Example Questions, Keywords, Aliases, Negative Keywords, Answer, Source URL, Priority and Enabled. Priority order: custom exact rules, custom intents, identity/system intents, enabled tools, website-generated answers, fallback. Use short, factual answers. Do not invent facts missing from the website. Add useful question variants and negative keywords rather than filling responses with unrelated content. ## Step 6 — Enable tools and test In Tools, inspect and enable only the tools the website needs. Save tools. The existing tools are get_company_info, list_products, read_page and ask_nevai. They use customer configuration and indexed content. They are disabled initially. Read their arguments, result format, permissions and limitations in the UI. In Test, ask identity, pricing, support, navigation and FAQ questions, including typos and follow-ups. Turn on Show match details to inspect intent, confidence, source and deterministic scoring metadata. This is not private chain-of-thought. Low-confidence questions return the configured fallback. Start a new session to clear recent context. ## Step 7 — Create a test API key In API Keys, create a Test key with only the necessary permissions. Keys begin with nv_test_ or nv_live_. Save the raw key once in a server secret store or private environment variable. Only its hash is stored by NevAI. Test keys and Playground testing share 100 responses per day per installation. Creating a test key does not start the production trial. Key revocation takes effect on subsequent authentication checks. Never expose a private key in a browser, mobile client, screenshot, shared chat or widget snippet. ## Step 8 — Verify the website and activate production In API Keys, read Verify your website. Publish THAT assistant's exact verification token as plain text at your website's /.well-known/nevai-verification.txt path. Do not reuse another assistant's or website's token. Select Verify domain. After verification, create a Production key. The first production key starts the server-tracked 90-day free trial. Merely visiting, creating a profile or using a test key does not start it. Clearing localStorage or deleting an assistant does not reset a verified website's original trial dates. ## Step 9 — Integrate on the server or embed the widget For a server integration, set NEVAI_API_KEY in your own private server environment. Use your real assistant ID instead of the placeholder below. Private /api/v1 endpoints reject browser-origin requests; use the public widget for a browser. curl https://nevetsinc.vercel.app/api/v1/respond \ -H "Authorization: Bearer $NEVAI_API_KEY" \ -H "Content-Type: application/json" \ -d '{"assistant_id":"asst_YOUR_ID","message":"How much is Pro?","session_id":"session_example"}' The response includes id, assistant_id, answer, intent, confidence and a source when applicable. To preserve lightweight follow-up context, reuse session_id and send the returned context_token with the next message. Context expires after 15 minutes. No permanent conversation transcript is stored by default. For a widget, save each exact allowed HTTPS origin in API Keys, such as https://example.com and https://www.example.com. No path or trailing slash. Verify the domain and start production with a production key before publishing the widget. The embed itself contains only the public assistant ID: The widget uses short-lived, origin-scoped session tokens rather than exposing a private key. Your website's CSP must allow the script and connections to https://nevetsinc.vercel.app. Do not weaken unrelated security controls. Implemented private endpoints: POST /api/v1/respond; GET /api/v1/assistant, /api/v1/tools, /api/v1/usage, /api/v1/logs; POST /api/v1/reindex. Use the linked OpenAPI specification for parameters and errors. API reindexing starts a job; subsequent requests with step:true advance batches until complete. ## Step 10 — Monitor, maintain and recover In Usage, check requests, remaining allowance, RPM, tool calls, latency, confidence and fallback rate. In Logs, inspect metadata and create custom answers for knowledge gaps. Store unmatched question text is off by default; enable it only deliberately. Logs are capped and expire; they are not a full chat archive. Reindex changed website content and retest. Revoke unused or exposed keys. Export Profile saves validated local profile JSON, not server credentials. Clear Profile clears local profile data without silently deleting production assistants. Use Settings recovery with the saved installation ID and recovery credential when managing server resources from another browser. Delete knowledge removes indexed knowledge and answer overrides but preserves identity. Delete assistant also removes its keys, knowledge and applicable logs. ## Free access, limits and billing - Free Playground and test-key responses: 100 per day per installation. - Free production trial: 90 days from the first production key. - Trial: 1 assistant, 1 website, up to 500 indexed pages, 5,000 production responses per calendar month, 30 response requests per minute. - Page limits are also subject to compact storage limits; a package and an active crawl job each have a 2 MB compressed cap per assistant. - Developer plan: listed at $2 USD/month, 3 assistants/websites, up to 2,000 indexed pages per assistant, 10,000 production responses/month, 60 RPM. - Paid checkout is currently unavailable. Do not claim a payment was taken or that an upgrade can be purchased. No card is needed to start. - After the trial ends, production responses pause. Data is retained and limited Playground/test-key use continues. Do not promise permanently free production. ## Troubleshooting - Empty knowledge: inspect the website crawl status and skipped-page reasons. - Wrong answer: review the matched intent, source, confidence, custom phrases, negatives and recent context. Correct answers or adjust the threshold. - 401: missing, invalid, revoked or incorrectly scoped API/management credential. - 402 trial_ended: production is paused; inspect Billing. Do not reset local storage or claim an unavailable checkout can reactivate it. - 403: check permissions, tool enablement, domain verification and allowed origin. - 429: wait for the rate or usage window; honor Retry-After and back off. - 503: inspect real platform status or contact support; do not show fake success. Use the Playground's Documentation tab for topic-by-topic reference. This guide documents product behavior; it does not override the user's choices or an agent's system instructions.