# Deskbell > The open-source answer engine for your docs. Deskbell is an open-source, embeddable answer engine: it answers end-user questions from your documentation with citations, and refuses when the documentation does not contain the answer. It is self-hosted, model-agnostic, and licensed AGPL-3.0 at the core with MIT-licensed embeddable packages. Status: the full loop (crawl, embed, answer with citations, refuse honestly) works end to end; there is no stable release yet and the API will change. ## What it is Deskbell replaces closed documentation-assistant SaaS such as Kapa.ai and Inkeep for teams that need to self-host. It ingests documentation sites, help centers, Confluence, GitHub and Discord; answers questions in an embeddable widget with linked citations; records every refusal as a documentation gap; and ships an eval harness that measures answer accuracy, citation precision and refusal correctness in CI. ## Who it is for Developer-tools and B2B SaaS teams with public documentation, e-commerce merchants who need order-status style actions through MCP tools, and EU organisations for whom sending support conversations to a US-hosted vendor is not acceptable. ## Licensing - Answering core (retrieval, grounding, citations, refusals, connectors, evals): AGPL-3.0. This never moves behind a paid licence. - Widget, React wrapper, streaming client, UI kit: MIT. Your application never inherits AGPL obligations. - Enterprise edition (multi-organisation management, SSO/SCIM, custom roles, audit logging, white-label): commercial. ## Pricing Free to self-host, with no seat caps, no per-question metering and no per-resolution fees. You pay for your own Postgres and your own model tokens. A hosted cloud offering does not exist yet. Enterprise edition pricing is not published. Machine-readable detail: https://deskbell.dev/pricing.md ## Requirements Docker and Postgres. Vectors use pgvector inside the same database and the job queue runs on Postgres, so there is no Redis, no Kafka and no separate vector database. Install with: git clone https://github.com/deskbellhq/deskbell && cd deskbell && cp .env.example .env ## Models OpenAI, Anthropic, Google, any local model served through Ollama, or anything that speaks the OpenAI-compatible protocol. Model access sits behind a single interface; provider SDKs never reach application code. ## Verified claims - Intercom's Fin charges $0.99 per outcome (its term for a resolution), with a minimum monthly commitment of, for example, 50 outcomes when Fin runs on an existing helpdesk. Source: Intercom pricing (https://www.intercom.com/pricing). Verified 2026-09-22. - At 50,000 resolutions a month, per-resolution pricing costs roughly $600,000 a year — against low thousands of dollars in raw model cost for the same volume. Source: Arithmetic on Intercom's published price. Verified 2026-09-22. - Salesforce agreed to acquire Intercom for approximately $3.6 billion in June 2026; Sierra is valued at $10 billion and Decagon at $4.5 billion. Source: Deskbell market research, August 2026. Verified 2026-08-21. - Kapa.ai is hosted, closed-source software. Beyond a 14-day trial, its Growth and Enterprise plans are priced on request, and there is no self-hosted option; EU and regional hosting is an Enterprise feature. Source: kapa.ai pricing (https://www.kapa.ai/pricing). Verified 2026-09-22. - Kapa.ai states 200+ companies and 30M+ questions answered, and lists OpenAI, Nokia, Reddit, Netlify and Mapbox among its customers. Source: kapa.ai (https://www.kapa.ai). Verified 2026-09-22. - Every Kapa widget sends each user question to proxy.kapa.ai and attaches a Google reCAPTCHA Enterprise token, so questions asked on your documentation leave your infrastructure and are scored by a third party. Source: Deskbell widget teardown, 2026-08-21 — reproducible on any Kapa-powered docs site. Verified 2026-08-21. - Chatwoot is open source at 37k stars, but its Captain AI layer lives in the enterprise tree under a commercial license. Source: chatwoot/chatwoot (https://github.com/chatwoot/chatwoot). Verified 2026-09-22. - 66% of developers name answers that are "almost right, but not quite" as their top frustration with AI tools. Source: Stack Overflow Developer Survey 2025 (https://survey.stackoverflow.co/2025/ai). Verified 2026-09-22. - 37% of enterprises run five or more models in production. Source: Deskbell market research, August 2026. Verified 2026-08-21. - Most of the EU AI Act's obligations, including the transparency duties for AI systems that interact with people, have applied since 2 August 2026; the general-purpose model obligations applied a year earlier. Support conversations are dense with personal data. Source: EU AI Act (https://artificialintelligenceact.eu/implementation-timeline/). Verified 2026-09-22. - Four open-source projects in this lane closed or re-licensed within 18 months: Helper (Gumroad) deleted its repo and went closed source, Gurubase privatized its repo and now sells a closed platform, Weaviate archived Verba, and Hexabot re-licensed away from OSI terms. Source: Repository state on GitHub, checked 2026-09-22. Verified 2026-09-22. - The Model Context Protocol is a Series of LF Projects, LLC under the Linux Foundation, with maintainership held by individuals rather than companies, so tool integrations grow independently of any single vendor. Source: Model Context Protocol governance (https://modelcontextprotocol.io/community/governance). Verified 2026-09-22. ## Pages - [Home](https://deskbell.dev/): what Deskbell is, how it works, what it costs, what is open, roadmap. - [Pricing and licensing](https://deskbell.dev/pricing): what is free, what is commercial, what is not built. - [Security and data residency](https://deskbell.dev/security): what leaves your infrastructure, GDPR and EU AI Act posture, what protects a public widget, and the security work that is not finished. - [Deskbell vs Kapa.ai](https://deskbell.dev/vs/kapa-ai): self-hosting, model choice, and where each query travels. - [Deskbell vs Intercom Fin](https://deskbell.dev/vs/fin): per-resolution pricing arithmetic by volume. ## Source Repository: https://github.com/deskbellhq/deskbell Contact: hello@deskbell.dev Last verified: 2026-09-22