kapa.ai is an ingestion and retrieval platform that turns a technical company's unstructured knowledge into cited answers for AI agents and for people. It builds one knowledge base from more than 20 source types, including documentation sites, GitHub code, issues and pull requests, Zendesk tickets, Salesforce Knowledge, Confluence, Notion, Jira, Slack, Discord, Discourse, Stack Overflow, OpenAPI specs, Google Drive, S3, PDFs and YouTube transcripts, and keeps it in sync as that content changes. Agentic retrieval searches it at query time and returns ranked passages with their source URLs, and every generated answer carries inline citations. When its sources do not cover a question, kapa.ai is designed to say it does not know rather than guess. The company, kapa.ai, Inc., went through Y Combinator in Summer 2023.
There are two ways to use it. Developers hand kapa.ai to their own agents as a search tool, through the Retrieval API or a hosted MCP server that each project can create in a few clicks, and it works with LangChain, Mastra, n8n, the OpenAI Agents SDK, the Anthropic SDK and the Vercel AI SDK. Its Kapa AI node is a verified n8n node, and its website script supports WebMCP, a draft web standard that lets a site offer tools to visiting AI agents. Teams that do not want to build an agent can deploy pre-built ones on the same knowledge base: a website widget, Slack and Discord bots, a support form deflector, a Zendesk agent that drafts grounded replies, an internal assistant, and in-product agents built with its chat and agent SDKs. A hosted MCP server can be public for users of tools such as Cursor, Claude Code and VS Code, restricted to employees with Kapa accounts, or keyed for a company's own agents. Analytics show the questions users ask most and the gaps in the documentation behind them.
Pricing is usage-based and quote-only. The free trial runs 14 days with no credit card and covers one index, 5,000 crawled pages, 5,000 retrieval queries and 1,000 answers, with the Retrieval API, the hosted MCP server and the pre-built agents included. Growth is priced as a platform fee plus a committed monthly answer volume, with overages at the same rate, and retrieval-only usage is priced separately and below answers. Enterprise adds SSO, SCIM, role-based access control, data retention controls, EU and regional hosting, and premium support with an SLA. Neither paid plan publishes a price, and commercial plans carry a minimum annual commitment. Customer success and forward-deployed engineering are included on every plan, and qualifying open-source projects can apply for free access.
The trial does not include private or authenticated sources such as Confluence, Notion, Zendesk, Salesforce and Google Drive, continuous sync (trial sources refresh manually), zero data retention, PII masking, or email and shared Slack support, all of which start on Growth. SSO, SCIM, role-based access control, data retention controls and EU hosting are Enterprise features. When a trial limit is reached, requests pause until the monthly reset or an upgrade, and after 14 days the workspace pauses until a paid plan starts. Paid plans have no self-serve card signup, and the vendor scopes each setup with the customer.
kapa.ai is best for companies with a technical product, such as developer tools, software, semiconductors and hardware, that want their documentation, code and support history to answer developers and customers accurately, with citations, wherever they ask. Teams whose agents need live results from the open web rather than from their own sources should look at Exa AI. Teams that want company-wide search and agents across every internal app, for every department, should look at Atlassian Rovo. Teams that want an AI agent resolving general customer service conversations inside a helpdesk, billed per outcome, should look at Intercom Fin.
Current state Q4 2026: In June 2026 kapa.ai introduced Kapa for Agent Context, which packages its retrieval for agents that teams build themselves, and it runs a public MCP server over its own documentation. In October 2026 it released its n8n node on October 12 and WebMCP support on October 16. It uses models from OpenAI, Anthropic, Cohere and Voyage alongside its own retrieval, embedding and re-ranking models, with training opt-outs signed with each model provider. Its trust center lists a SOC 2 Type II report dated April 2026 and GDPR. Named customer stories report a 98% ticket deflection rate at monday.com, 88% of Raspberry Pi users self-serving, and more than 700,000 queries processed at Nordic Semiconductor.
