Private AI (now Limina) alternative for LLM apps

If you bookmarked private-ai.com a while ago, it now redirects to getlimina.ai. On 5 March 2026 the Toronto company announced it had rebranded as Limina; it says products, customers and operations are unchanged. So if you're searching for a Private AI alternative, you're comparing against Limina.

Maskflare wrote this page, so weigh it accordingly. Limina and Maskflare overlap less than the search results suggest. One is a de-identification engine for regulated documents and datasets. The other is a privacy layer that sits in the path of every AI request your apps, people and agents make.

What is Limina?

Limina sells data de-identification. According to its website as of October 2026, it detects 50+ entity types across PII, PHI and PCI in 52 languages, including text that switches language mid-sentence. Detection is model-based. It can redact, pseudonymize, generate synthetic replacements or generalize entities in medical records, call transcripts, chat logs and documents.

Deployment is a strong point. The engine ships as a container that runs in your VPC or on-premises, and there's also a SaaS version and marketplace listings on AWS, Snowflake and Azure. Its focus is regulated industries: pharma and life sciences, healthcare, financial services, contact centers and insurance. The pricing page lists a free Starter tier (75 API calls a day, 10 per minute), a Batch tier for large one-off jobs and a custom Enterprise tier.

What happened to PrivateGPT?

In May 2023 Private AI launched PrivateGPT, a privacy layer for ChatGPT: it replaced identifiers with markers like [NAME_1] before the prompt left your network, then swapped the original values back into the completion. As of October 2026 PrivateGPT doesn't appear on Limina's website and its old documentation pages return errors, so it appears to be retired. We couldn't find an official announcement either way; ask Limina if you depend on it.

Two things to keep straight. First, the redact-then-restore pattern PrivateGPT offered is exactly what Maskflare's gateway and browser protection do. Second, there's an unrelated open-source project called PrivateGPT by Zylon, an API layer for local models. It's not a de-identification tool.

When is Limina the better choice?

• You're de-identifying data at rest. Clinical notes for research, call transcripts for analytics, a dataset you want to use for model training. Batch de-identification with synthetic replacement is Limina's home ground.
• You need many languages. 52 languages is far more than Maskflare's name detection (English, German, French, Spanish, Italian, Albanian). If your contact center handles Portuguese, Arabic and Japanese, Limina is likely the better fit for names and free-text entities.
• Names, addresses and medical concepts in prose matter most. A trained model generally does better than rules on unstructured entities like those.
• You want synthetic data, not tokens. If downstream systems need realistic fake names rather than placeholders, Limina offers synthetic PII generation.

When is Maskflare the better choice?

• You need values restored in live answers. Maskflare masks values before they reach OpenAI, Anthropic, Gemini or six other providers, then restores them where the answer repeats them, including while the response streams. With a de-identification API you build that round trip yourself.
• You want a drop-in gateway. Change one base URL. No calls to a separate detection API, no rebuilding request bodies per provider.
• Your people use ChatGPT and Claude in a browser. The desktop agent and browser protection mask what employees type into ChatGPT, Claude, Gemini, DeepSeek and Perplexity on the web.
• Agents and MCP are in scope. Tool-call arguments and results are masked on the MCP route, and hooks cover Claude Code, Cursor and Codex CLI.
• Structured identifiers and secrets matter most. 47 country packs with checksum validation, 232 credential formats, and Exact Data Match against your own customer records.
• You want results you can explain. Detection is deterministic, so the same input always gives the same result, and the explain view shows which rule fired.

Limina vs Maskflare, feature by feature

How does the Maskflare round trip work?

Say a support agent asks your internal assistant: “Summarize the complaint from Ada Lovelace, IBAN CH93 0076 2011 6238 5295 7, about a CHF 2,400 charge.” The gateway replaces the name and IBAN with tokens such as mf_mask_170b... before the request leaves for the provider. The model writes its summary around the tokens. On the way back the gateway swaps the real values in, so the agent reads a normal answer. The provider never saw the IBAN, and Maskflare doesn't log the prompt or the response.

from maskflare import MaskFlare

client = MaskFlare(api_key="mf_live_...", base_url="https://<your-gateway>")

result = client.mask("Refund Ada Lovelace, ada@example.com, CHF 2,400")
# result.output -> "Refund mf_mask_9c2e..., mf_mask_170b..., CHF 2,400"
# send result.output to any model, then restore the reply:
answer = client.unmask_text(model_reply)

The honest objections

“Limina supports 52 languages.” It does. If multilingual names in free text are your main risk, test both: Maskflare's named-entity layer finds names in free text, and our structured-identifier coverage across 47 countries goes well beyond generic entity types.

“Limina runs entirely in our VPC.” So can Maskflare, for enterprise customers: on-premises deployment is available. By default we're EU-hosted.

“We could use both.” Yes, and some workloads split naturally that way: batch de-identification of a training corpus in one tool, live masking of prompts and agent traffic in another. Just don't run two detectors over the same live request; you'll get double-masking and harder debugging.

“Model-based detection is more accurate.” On names and addresses in prose, often. On a Swiss AHV number with a valid check digit or an AWS secret key, rules are precise and predictable. Test both on your own data; neither vendor's numbers replace that.

What Maskflare doesn't do

• No synthetic data. If a research team needs a corpus where every name is replaced with a realistic fake one, Maskflare's tokens won't do.
• Test names on your own data. Person names are found by word lists, context and the named-entity layer, and structured identifiers cover 47 countries. Check unusual names from your own documents, and add a dictionary for known customers.
• No bulk document de-identification product. Maskflare scans documents (PDF, Office, HTML, email) in a sandbox and blocks or allows them by policy; it masks plain-text files in place. It doesn't rewrite a PDF with redactions.
• Certifications in progress. Maskflare's SOC 2 Type II and ISO 27001 audits are in progress. If your procurement needs a report today, ask us for the security documentation we share during evaluation.

If those are deal-breakers, Limina is a reasonable choice and we'd rather tell you now. If your problem is live prompts, employee AI use and agents, start with a trial and see how far deterministic detection gets you on your own traffic.