Transparent Pricing for Every Scale
Start with our free community engine in Observe Mode. Upgrade when you need machine-assisted policy learning, Presidio/GLiNER NLP, or enterprise compliance ledgers.
Developer Node
Essential AI Privacy WAF for indie hackers and small local swarms.
- Observe Mode & Enforce Mode
- Custom Regex Rule Engine
- Reversible Surrogate Tokenization
- Up to 1,000 Active Vault Tokens
- Up to 10 Custom Allowlist Entries
- Community Discord Support
Team & Agency
Machine-assisted WAF policy recommendations and Presidio NLP.
- Machine-Assisted Policy Recommendation Engine
- 1-Click Staged Promotion & Allowlists
- Presidio NLP Local Entity Recognition
- Up to 50,000 Active Vault Tokens
- Business Jargon False-Positive Shield
- Full Telemetry Prompt & Response Audit Search
- Down to 5-Minute Automated Token Pruning
Global Enterprise
Unlimited throughput, GLiNER NER, and regulatory compliance.
- Unlimited Active Vault Tokens
- GLiNER Zero-Shot Named Entity Recognition
- Multi-Tenant & Multi-Region Isolation
- Cryptographic Audit Ledger (ISO 27001 / EU AI Act)
- Air-Gapped Sovereign Node Support
- Custom Regex & NER Model Fine-Tuning
- Dedicated SRE Support & 99.99% SLA
Frequently Asked Questions
How does Observe Mode prevent prompt degradation?
Observe mode operates as a transparent audit proxy. Detections are logged and analyzed in the background without altering the text sent to the LLM. You only promote rules after reviewing machine suggestions.
How does TokenShield handle business jargon?
Our contextual filter analyzes technical stop words and token entropy. Phrases like "Compute Environment" or "500 K Q3" are automatically categorized as safe business jargon with a low threat score and proposed for allowlisting.
Can I deploy TokenShield entirely on-premises?
Yes. TokenShield can be deployed as an autonomous Docker container or Podman VM node connected to your private PostgreSQL instance. Zero telemetry ever leaves your perimeter.
Does surrogate tokenization affect LLM reasoning?
No. Surrogate tokens maintain grammatical structure and semantic placeholders, enabling the model to reason accurately over relationships before the response is de-tokenized back to original values.