Comparison
Deskbell vs Kapa.ai
Kapa.ai is a closed SaaS documentation assistant with per-question pricing quoted on request. Deskbell is an open-source alternative you run yourself, on any model, with the eval harness included. Kapa is far more mature; Deskbell is new. Here is the honest breakdown.
Last updated: 2026-09-22
Where Kapa wins
Maturity, and it isn't close
- It works in production today, for OpenAI, Nokia, Reddit, Netlify and Mapbox — some of the highest-traffic documentation on the internet.
- Years of retrieval tuning against real developer questions, which is the part that takes longest to get right.
- A managed service: no Postgres to run, no embeddings pipeline to babysit, no upgrade path to own.
- A polished widget with the small details — keyboard handling, mobile behaviour, streaming edge cases — already solved.
If your requirement is “working answer engine on our docs this quarter, budget available,” Kapa is the safer choice today. We would rather say that than have you find out after a migration.
Where Deskbell differs
Ownership, and what that unlocks
- It runs on your infrastructure. Conversations — which are dense with personal data — stay in your database, in your region.
- Any model, including a local one. Your compliance posture picks the provider, not your vendor's roadmap.
- Free to self-host, with no per-question meter, so cost doesn't scale with your documentation traffic.
- The eval harness ships with the product, so answer quality is something you measure rather than something you're told about.
None of that matters if the answers are bad, which is why the eval harness is in the box and why we publish our own numbers.
Data flow
Where a question actually goes when someone asks it on your docs
We opened a Kapa-powered documentation site and watched the network tab. This is not a criticism of the engineering — it is the standard architecture for a hosted widget — but it is worth knowing before a privacy review, because it is a different trust boundary than most teams assume when they embed something on their own domain.
Kapa widget, per question
- 1The question is posted to
proxy.kapa.ai, not to your servers. - 2A
x-recaptcha-enterprise-tokenis attached to every query, so each visitor to your documentation is also scored by Google. - 3The payload carries a persistent anonymous visitor id and the page URL the question was asked from.
- 4The answer streams back from Kapa's infrastructure.
Reproduce it yourself: open any Kapa-powered documentation site, open your browser's network tab, ask the widget a question, and look at the outbound request. Takes about thirty seconds.
Deskbell widget, per question
- 1The question is posted to
your-domain.com/api/v1/ask— an endpoint you operate. - 2Retrieval runs against your Postgres. No third-party scoring, no external bot service on your docs page.
- 3One outbound call goes to the model provider you configured — or none at all, if you run a local model.
- 4The transcript is stored in your database, under your retention policy.
The trade-off is real: you operate it, you patch it, and you own the abuse protection. Self-hosting moves the work, it doesn't delete it.
Side by side
Feature comparison
| Capability | Deskbell | Kapa.ai |
|---|---|---|
| Self-hostable | Yes — docker compose, Postgres only | No. Closed SaaS, no self-host option |
| License | AGPL-3.0 core, MIT widget and SDKs | Proprietary |
| Pricing | Free to self-host. No per-question metering | On request beyond a 14-day trial; the trial is capped in answers |
| Model choice | OpenAI, Anthropic, Google, or local via Ollama | Vendor-managed |
| Where queries go | Your server. Only the model call leaves — or nothing, with a local model | Proxied through proxy.kapa.ai, with a Google reCAPTCHA token attached |
| Citations in answers | Yes, linked to the source section | Yes, linked to the source section |
| Answer-quality evals | Eval harness ships in the repo and runs in CI | Internal, not shipped to customers |
| Docs-gap analytics | Unanswered questions ranked, in the free core | Analytics included in the product |
| Actions beyond answering | Any MCP server's tools, each off, allowed or needs-confirmation; writes run only after the visitor confirms a card; every call in an audit log | Not offered |
| Maturity | New, no stable release | Production-grade; lists OpenAI, Nokia, Reddit, Netlify and Mapbox as customers |
Kapa does not publish pricing, so that row reflects what is publicly stated rather than a quoted figure. Something inaccurate? Open an issue and we'll correct it.
Sources: kapa.ai pricing · kapa.ai · Deskbell widget teardown, 2026-08-21 — reproducible on any Kapa-powered docs site
Choosing
Which one is right for you
Choose Kapa.ai if…
- You need something live and dependable this quarter.
- Per-question pricing fits your budget and traffic.
- You have no appetite for running another service.
- Sending documentation questions to a US-hosted vendor is acceptable to your legal team.
Choose Deskbell if…
- Self-hosting or EU data residency is a hard requirement.
- You want to pick the model, including a local one.
- Your documentation traffic makes per-question pricing painful.
- You want to measure answer quality yourself rather than take a vendor's word for it.
- You can live without a stable release, and would rather shape the project than wait for it.
Questions
Common questions
Is Deskbell a drop-in replacement for Kapa.ai?
Not today. Deskbell is new: the ingestion, grounding, citation and refusal loop works end to end, but there is no stable release and the API will change. Kapa is a mature product serving large developer platforms. Evaluate Deskbell if self-hosting or model choice is a hard requirement.
Can Kapa.ai be self-hosted?
No. Kapa is a closed SaaS product. The widget runs on your documentation site, but ingestion, retrieval and generation happen on Kapa's infrastructure, and there is no on-premise or self-hosted distribution. That is the structural difference between the two products, not a feature gap.
How much does Kapa.ai cost?
Kapa publishes no price beyond a 14-day trial; Growth and Enterprise are priced on request, and the trial is capped in answers and retrieval queries, so usage is what is being sold. You cannot model the cost before talking to sales. Deskbell self-hosted costs the Postgres instance plus whatever tokens your chosen model provider charges.
Do my users' questions leave my infrastructure with Kapa?
Yes. Each widget query is posted to proxy.kapa.ai together with a Google reCAPTCHA Enterprise token and an anonymous persistent visitor id. This is a reasonable architecture for a hosted SaaS, but it is a different trust boundary than self-hosting, and it matters under GDPR review.
Which one should I pick right now?
If you need something dependable this quarter and per-question pricing is acceptable, pick Kapa. If self-hosting, EU data residency, model choice or auditable answer quality are requirements, Deskbell is built for that — but you would be adopting a project without a stable release and should plan accordingly.
Try it against your own docs
Point it at your documentation and see how it answers before you commit to anything. It's one command, and nothing leaves your machine.