Conversation inboxEvery answer stays reviewable
DW
Duplicate subscription chargeResolved
TO
API limits for our launchOpen
LS
Needs a billing specialistHandoff
Build a support agent from the knowledge you already own.
Point ChatGrain at your website. It reads the useful pages, remembers where every fact came from, and gives your customers a fast answer or an honest “I don't know.”
Tested on a real JavaScript-rendered sitemapLocal benchmark, 92 public URLs
92sitemap URLs found
88useful pages retained
88kcharacters extracted
58slocal crawl time
From a URL to a useful answer.
The training path is inspectable at every stage. You can see what was discovered, what was retained, and what supported each response.
Discover without hiding the work.
ChatGrain respects robots.txt, rejects private-network URLs, detects redirect loops and soft 404s, and keeps the queue within your page limit.
01sitemap.xml
02docs/getting-started
03help/refunds
The work starts after the first answer.
Review conversations, improve weak responses, adapt the widget, and connect the agent to real support work.
Review conversations
Answer qualityEvidence, not vanity metrics
91%grounded resolution
Answered with sources+12% this month
Your brand, not oursTry a widget color
Actions and handoffFinish work or involve a person
Customer asks Agent checks evidence Human gets context
Handoff requestedTranscript and matched sources attached
Sources stay attached from document to answer
Automation should make support more human, not less.
ChatGrain handles the repeated questions and preserves the context for the ones that need judgment. Your team sees the source, conversation, feedback, and visitor details in one place.
Verified repliesEvery factual answer can point back to evidence.Context-rich handoffPeople start where the agent stopped.Continuous improvementFeedback and pinned answers close knowledge gaps.
A working platform, not a landing-page promise.
These capabilities exist in the current application today. The roadmap is kept separate so “planned” never looks like “shipped.”
Website and sitemap training
Static pages and JavaScript-rendered sites, with robots rules, soft-404 detection, deduplication, and bounded crawling.
Available nowGrounded answers
Hybrid retrieval, source citations, pinned answers, and an evidence threshold that refuses unsupported questions.
Available nowConversation inbox
Persistent visitor history, unread operator replies, tickets, attachments, feedback, and a clear path from automation to human review.
Available nowAgent actions
Define JSON actions, collect leads, and connect the support experience to the workflows your team already owns.
Available nowKeep the support brain on your own infrastructure.
Run the crawler, worker, retrieval, dashboard, and widget on infrastructure you control. Choose local models now and add hosted providers later without replacing the product.
PostgreSQL + pgvector Signed public widget access Provider adapters Runs on your own server
Open the dashboard PS npm install
PS npm run services:up
PS npm run db:migrate
PS npm run dev
ready web and worker running
ready PostgreSQL connected
ready local embeddings availableOne licence. Honest infrastructure costs.
There is no pretend pricing for a billing system that does not exist yet. Run the complete current product on your own infrastructure now; a managed service can come later.
Self-hosted
Your servers your data
You pay only for the infrastructure and models you choose. Nothing leaves your deployment.
- Unlimited agents in your own deployment
- Website, text, and file sources
- Embeddable branded widget
- Local embeddings with Ollama Cloud generation
- Conversations, leads, feedback, and analytics
- Image understanding and saved voice messages
- PostgreSQL and pgvector storage
Managed ChatGrain
Later
Hosted workers, managed upgrades, backups, team billing, and production support.
Managed browser workersEncrypted backupsUsage metering and billingRegional deployment options
See what remainsQuestions worth answering clearly.
No invented customer counts, fake testimonials, or feature badges for work that has not shipped.
Test the product yourselfYes. It uses a fast HTTP path first, then opens a controlled browser only when a page returns an empty JavaScript shell. That keeps ordinary sites fast without losing content from modern React and Next.js websites.
Every response begins with retrieved passages from your sources. Weak evidence produces a refusal instead of a confident guess, citations show the supporting page, and pinned answers can override retrieval for critical questions.
The crawler, worker, retrieval pipeline, dashboard, and widget all run on infrastructure you control, and your content never leaves it. You pay for your own hosting or model hardware, and ChatGrain does not require a paid AI API.
The crawler detects a site name, icon, logo, and dominant color. You can then adjust the agent name, welcome message, theme color, launcher, and allowed domains before publishing.
The current build includes Clerk authentication, isolated workspaces, administrator controls, audit logs, signed widget tokens, URL safety checks, rate limits, retention, and Sentry instrumentation. A serious deployment still needs encrypted backups, managed secrets, a reviewed privacy policy, and recovery testing.
Your documentation is already doing the hard part.
Give it a support interface your customers can actually use.