Kastr vs TalkingPoints: translation depth against district infrastructure
TalkingPoints is a non-profit whose entire reason for existing is language access for families schools were failing to reach. On the specific question of translation quality in low-resource languages, they are probably better than us, and this page starts there because pretending otherwise on a page about translation would be absurd.
| Aspect | TalkingPoints | Kastr |
|---|---|---|
| Approach | Human-in-the-loop review layered on machine translation | DeepL machine translation, no human review |
| Low-resource language quality | Likely better | Only as good as the engine |
| Idiom and school-specific phrasing | A reviewer catches it | Nothing catches it |
| Custom district glossary | Human review substitutes | Not built — no glossary of any kind |
| Preview before sending | Depends on workflow | Render the draft in several languages first |
| Send a test to yourself only | Not published | Yes, prefixed [TEST] |
| Behaviour when translation fails | Not published | Honest passthrough, target language still recorded |
| Cost of the translation layer | Free — grant funded | In the per-student rate |
| District admin surface, API, audit | Narrower | 28 endpoints, hash-chained audit |
| Long-term funding certainty | Philanthropic | Pre-launch commercial — also unproven |
The last row cuts both ways and we have marked ourselves amber on it deliberately. A grant-funded non-profit and a pre-launch startup are both organisations a district should ask hard questions about.
Why we will not print a language count
Most vendors in this category advertise a number of supported languages. This site used to carry one and we removed it, because we could not find a constant in the code that justified it. That is an embarrassing thing to publish and it is also the most useful thing on this page.
What is actually true, checkable in a trial: translation runs on DeepL's v2 API, whose real target set is around thirty languages. The composer surfaces nine for preview. A family profile offers ten. The ParentSquare importer maps thirteen inbound language codes. Those are four different numbers describing four different surfaces, and none of them is a headline marketing figure.
So the honest description is "DeepL-powered translation, roughly thirty target languages, nine exposed in the composer today." If the specific languages your families read matter — and they do — the only responsible answer is to check yours against DeepL's published list before signing anything, with any vendor. A district with a significant Hmong, Marshallese, Karen or Somali population should be extremely sceptical of every machine translation claim in this market, including ours.
What we do not have. No custom or district glossary — the DeepL glossary API is not called anywhere in our code. No "see original" footer on translated messages. Both appear on competitor feature lists and neither exists here. If a district-specific term must render consistently every time, we cannot guarantee it and a human-review service can.
The workflow difference, which is the real trade
Machine translation is fast and free at the margin; human review is accurate and has a queue. That difference shows up in workflow rather than in a feature table.
Our answer to the accuracy gap is to make the machine output visible before it leaves. The composer renders your draft in several target languages while you are still writing it, so a sentence that translates badly can be rewritten rather than discovered afterwards. Send-test-to-me delivers a copy to you alone, prefixed [TEST], which is the cheapest way to catch a mangled merge before twelve thousand households see it. If a bilingual staff member reviews that preview, you have effectively assembled a human-in-the-loop workflow out of your own people.
Two engineering details that matter more than they sound. Translations are cached globally on a hash of source language, target language and text, so the same string is never paid for twice — and the cache holds no index of who sent what, so a row is retrievable only by someone who already has the exact source text. And when the translation API has no key or fails, we pass the original text through and still record the target language, rather than silently reporting a translation that never happened. That second one is a small thing that keeps your language analytics honest.
Funding models, argued in both directions
TalkingPoints is a 501(c)(3) and free to schools. That is a real advantage: no procurement, no budget line, no renewal negotiation, and a mission that will not be redirected by an acquirer. For a district with genuinely no budget and a serious language-access problem, they are the right call and we would say so on a call.
The fair counter-argument is that philanthropic funding is a different kind of uncertainty rather than an absence of it. Grant cycles end, priorities move, and a district has no contractual lever when they do. There is no term, no price lock and no exit right, because there is no contract.
The fair counter-counter-argument is that we are a pre-launch company with no customers, no revenue history and no SOC 2 audit, so we are hardly a picture of institutional permanence either. What we offer instead is written commitments: §3.2 fixes your rate for 36 months and caps year four at the lesser of CPI-U or five per cent; §7.1 gives you a right to a complete machine-readable export at any time, without notice or fee; §11.2 lets you leave within 90 days with a prorated refund if we are acquired or materially change our data terms. Those clauses are worth exactly as much as the company that signs them, which is why we publish them rather than describe them.
Who should buy which
TalkingPoints, if language access is the central problem, your budget is zero or close to it, and your requirements are teacher-to-family messaging rather than district notification infrastructure.
Kastr, if you need district-wide delivery with a voice channel and SMS failover, an audit trail that survives a records request, an API your team can script against, and one published price that cannot move for three years — and if machine translation with a preview step is sufficient for your language mix.
Both, which is more common than either vendor's marketing would suggest. Districts run TalkingPoints for classroom-level multilingual conversation and a platform underneath it for official notification. If you do that, decide which one families should install and be consistent about it, because the failure mode here is not cost, it is two apps and neither read.
Questions people actually ask
How many languages does Kastr translate into?
We deliberately do not publish a headline number. Translation runs on DeepL, whose target set is around thirty languages; the composer exposes nine for preview today and the family profile offers ten. Check your community's specific languages against DeepL's published list before signing with us or anyone else.
Is machine translation good enough for school-to-home messages?
For routine notices in well-supported languages, generally yes. For nuanced or sensitive communication, or for low-resource languages, a human-reviewed service will usually do better. The mitigation we offer is visibility: preview the translated draft before it sends and have a bilingual staff member read it.
Can we preview a translation before it sends?
Yes. The composer renders your draft in several target languages while you write, and send-test-to-me delivers a [TEST]-prefixed copy to you alone. Both exist specifically because machine translation should be inspected rather than trusted.
Is TalkingPoints free for districts?
Yes. TalkingPoints is a non-profit and free to schools and teachers, funded philanthropically. The trade is a narrower district-administration surface and no contractual commitments — no term, no price lock, no exit right, because there is no contract to hold them in.
Which is better for a district with many low-incidence languages?
Probably TalkingPoints, on translation quality. Machine translation degrades most in exactly the languages a low-incidence population speaks, and a human-review layer is the only reliable mitigation. If you also need district notification infrastructure, run both rather than compromising on either.
One price. Every feature. Locked for three years.
$3.50 per student per year under 5,000 students. No tiers, no add-on modules, no per-message fees. Published on the site because you should not have to book a call to learn a price.