| What it does | Catch your AI agents' mistakes, block the dangerous ones
| On-device moderation that flags hateful, abusive and threatening text in 23 European languages, before the message is sent.
|
|---|
| Stage | Startup
| MVP
|
|---|
| Founded | 2026
| Not given
|
|---|
| Based in | Not given
| NL
|
|---|
| Pricing model | Not given
| Freemium
|
|---|
| Pricing | Free: $0/month with 5,000 conversations/month, Langfuse or LangSmith integration, problems and Slack alerts, rules that block costly actions, no credit card required. Starter: $49/month for 25,000 conversations/month plus email support. Growth: $149/month for 100,000 conversations/month plus priorit
| Free up to 100k monthly active devices with unlimited inference. Custom licenses available.
|
|---|
| Key features | - Monitor AI agent conversations without code by connecting Langfuse or LangSmith
- Judge conversations on six criteria: accuracy, goal alignment, decision quality, completeness, helpfulness, and factuality
- Create code-based rules that block risky actions before they run
- Group failures into problems with suggested rules and failure counts
- Send Slack alerts for serious mistakes and reply checks immediately
- Add to agents with one line of Python code for OpenAI Agents SDK, LangChain and LangGraph
- Mask sensitive data like email addresses, phone numbers and card numbers before sending to Kalytera
- Rules continue running offline if Kalytera goes down
- Correct judge verdicts by marking conversations as passed or failed
| - Three separate scores: hateful, abusive, threat
- Ten target signals under the hate score: race, color, religion, descent, national or ethnic origin, sexual orientation, gender, disability, age, other
- Runs fully on device with no server calls
- Supports 23 European languages from one model file
- 80MB model size optimized for iOS, Android, and browser
- 0.847 macro F1 accuracy on Multilingual HateCheck
- Warn before message is sent
- Comment triage without infrastructure
|
|---|
| What makes it different | - Rules are code-based, not AI, so they cannot be overridden or argued with
- No code setup using Langfuse or LangSmith connections for immediate visibility
- Rules continue protecting agents even when Kalytera is offline
- Catches exact mistakes from real AI agent incidents like database deletes, account takeovers and crypto theft
- Customer data stays on your servers; Kalytera only reads conversations, never writes or stores sensitive data
| - Multilingual support across 23 European languages in a single 80MB model
- Complete privacy with no data leaving the device
- Free access up to 100k monthly active devices with unlimited inference
|
|---|
| Community votes | 0 | 0 |
|---|
| DR (Domain Rating by Ahrefs) | 1 ● no change since last check | 37 ● no change since last check |
|---|
| Trust Flow (Majestic) | 0 | 20 |
|---|
| Citation Flow (Majestic) | 14 | 39 |
|---|
| Referring domains (Majestic) | 28 | 616 |
|---|
Toxic