BYOK vs Managed Keys: What AI Security Teams Actually Choose

ekmire's free tier runs in two modes. Which one you need depends on a single question: can your code and request payloads touch a third-party LLM?

If yes, ekmire inference is the right starting point. If no — because you are in a regulated industry, running air-gapped, or your organization has a vendor policy about external data processing — BYOK is what you need.

This is not a tier question. Both modes are free. Both give you every ekmire detection feature. The choice is entirely about where AI analysis traffic goes when ekmire needs to evaluate a payload semantically.

What ekmire inference gives you

The default free tier routes the semantic analyzer through Groq-hosted Llama 3.3 70B. ekmire handles the LLM routing — you provide no API key, pay nothing for inference calls, and start catching prompt injection and AI-targeted attacks without a provider account.

The limit is 2,000 events per month. For a solo developer or a small project in early stages, that is typically enough headroom. The deterministic engines — Build Guard's regex and AST analysis, the proxy's rule matching — have no event cap and run entirely offline. The event count only affects cloud-logged detections and semantic analysis calls.

What ekmire can see in this mode: the payload content evaluated by the semantic analyzer, because that payload routes through ekmire's LLM endpoint. The raw content is not retained beyond the analysis call, but if your security posture requires that no third party ever sees your request payloads, this mode is the wrong fit.

What BYOK gives you

BYOK mode connects the semantic analyzer directly to your own LLM provider account. You supply the key; ekmire routes the analysis call from the proxy or CLI to your provider without the traffic touching ekmire's infrastructure.

The practical differences:

• 10,000 events per month (vs. 2,000 on ekmire inference) — not because BYOK is a higher product tier, but because ekmire is not absorbing LLM costs
• Unlimited dev-audit calls — ekmire dev-audit has no cap in BYOK mode
• Direct provider billing — inference costs appear on your OpenAI, Anthropic, or Groq invoice
• Zero ekmire LLM access — payloads travel from your running proxy to your provider, bypassing ekmire's routing layer entirely

The trade-off: setup takes two minutes instead of zero. You configure your key in the dashboard once. It is stored AES-256-GCM encrypted at rest, decrypted only in-memory during routing. The key is never logged, never transmitted to ekmire in plaintext, and can be deleted on demand from the Settings screen.

Supported providers and how to switch

BYOK currently supports four provider types:

• Groq — fastest option if you already have a Groq account; the same model ekmire inference uses by default
• OpenAI — GPT-4o or GPT-4o-mini for semantic scoring
• Anthropic — Claude Haiku or Claude Sonnet
• Ollama / OpenAI-compatible — any locally hosted model via an OpenAI-compatible endpoint; this is the path for air-gapped deployments where no outbound LLM call is acceptable

Switching providers does not require a restart. In the dashboard: Projects → your project → LLM provider. Select provider, paste key, save. The change takes effect on the next semantic analysis call. You can switch back to ekmire inference at any time by clearing the key.

# The CLI inherits the provider configured in your project dashboard.
# For local testing, override per-call with env vars:
EKMIRE_LLM_PROVIDER=ollama EKMIRE_LLM_BASE_URL=http://localhost:11434 ekmire dev-audit src/

The proxy picks up provider changes without a restart — the same hot-reload mechanism used for rule bundles applies. In-flight requests finish with the previous provider; new requests use the updated one. The two-line WAF setup guide covers how the proxy runs and what hot-reload looks like in a Docker Compose deployment.

Who actually needs BYOK

Three situations where BYOK is not optional:

Regulated industries. Healthcare, finance, and legal teams operating under HIPAA, DPDP, PCI-DSS, or equivalent frameworks often have explicit restrictions on sending sensitive data to third-party AI services not covered by a BAA or data processing agreement. ekmire's LLM endpoint is not party to those agreements. Your own OpenAI or Anthropic account — with a DPA signed — is. In 2026, this is increasingly enforced policy, not just a recommendation. Sovereignty is now the enterprise AI operating model: McKinsey, the WEF, and MIT Technology Review all published major analyses this year on AI data routing as a first-tier governance decision.

Existing enterprise provider relationships. If your org already has an enterprise OpenAI or Anthropic contract with volume discounts, centralised billing, and audit logging requirements, routing ekmire's AI calls through that account keeps everything in one audit trail. Your security team's LLM review already covers it.

Air-gapped or on-premises deployments. If your application cannot make outbound calls to ekmire's API domain at all — common in defence contracting environments and certain government infrastructure — Ollama running locally is the only path. BYOK with an Ollama endpoint gives you full semantic analysis with zero external network dependency. The proxy, the CLI, and the LLM all stay inside your perimeter.

What BYOK does not change

This is the part that surprises most people who ask about it.

The deterministic detection layer is entirely unaffected by your LLM provider choice. Build Guard's regex engine, the AST engine, the proxy's rule matching against the Ed25519-signed bundle — none of that uses the LLM. It runs offline, cached locally, and costs nothing per call regardless of tier or provider configuration.

BYOK is exclusively about the semantic analysis path: the second-pass LLM evaluation that fires on payloads that passed rule checks but scored above the log threshold, and the ekmire dev-audit command that reviews code for logic-level vulnerabilities the regex engine cannot reach. The other 90% of ekmire's detection surface is identical across both modes.

The threat feed is also unchanged. Ed25519-signed rule bundles are published and hot-reloaded the same way regardless of your provider. Your provider choice has no effect on rule coverage — BYOK is not a different product, just a different routing path for one of the three detection layers.

Three things to take away

1. Both modes are actually free. BYOK costs nothing to ekmire. If you are on Groq's free tier, BYOK on Groq can stay operationally free for a long time depending on your event volume. The 10,000 event cap reflects what you are not paying for, not a feature gate.
2. Make the data sovereignty call on day one. If you are building something that will eventually need DPDP, HIPAA, or enterprise security policy compliance, start on BYOK. Switching later takes 90 seconds in the dashboard — but it is extra work you do not need to schedule.
3. Local inference is production-viable. Ollama running Llama 3.1 8B is a legitimate BYOK option for solo developers who want air-gapped semantic analysis at near-zero cost. Latency is higher than Groq — typically 3–8 seconds vs. 600ms — but the proxy's 1,500ms semantic timeout means slow responses fall back to pass rather than holding a request indefinitely. Real requests are never blocked by a slow local model.

The BYOK setup guide covers provider-specific configuration, key rotation, and the Ollama endpoint format for local inference. The pricing page shows the full event limit comparison — BYOK is a free-tier feature, not an upgrade path.