

Portkey is one of the best production LLM gateways going — 1,600+ models behind one API, reliability routing, deep observability, guardrails, and virtual keys. But those virtual keys sit in Portkey’s encrypted vault, and Portkey is the process that holds and uses the real provider key to reach the model. KnoxCall attacks a different layer: the provider key is injected at the egress wire so it never enters your workload, and KnoxCall custodially rotates the underlying vendor child key — not just a store entry. Where Portkey wins on model coverage and observability, we say so.
Credential Exposure (the core difference)
Scope: third-party outbound bearer keys (OpenAI, Anthropic, Stripe, Twilio). Portkey’s job is to hold and use the provider key to reach the model, and it does that well — but that means the live key sits in Portkey’s store and its serving process.
| Feature | KnoxCall | Portkey |
|---|---|---|
Provider key never enters the process that reaches the model Egress wire injection: KnoxCall attaches the real key at the outbound hop, your workload never holds it | ✓
Injected at egress; no value-GET path | ✗
Virtual key maps to the real key stored in Portkey’s vault; Portkey’s process uses it to call the model |
Rotates the underlying VENDOR key, not just a store entry Custodial rotation mints / verifies / deletes the provider’s own child key on a lease schedule | ✓
Rotates the real vendor key itself | ~
Virtual keys ease rotation, but you still rotate at the provider yourself; the stored key is what you set |
Guards / redacts request & response CONTENT Filter, fix or block prompts and completions in flight | ✓
Streaming PII redaction + prompt firewall (Pro+) | ✓
60+ guardrails, its flagship strength |
DPoP-bound short-lived token in the workload (not a long-lived key) What the process actually holds, sender-constrained and revocable | ✓
Scoped, DPoP-bindable KnoxCall token | ~
Workload holds a Portkey API key / virtual key; not DPoP sender-bound |
Workload identity federation (OIDC token exchange, RFC 8693) Swap a workload’s OIDC identity for a short-lived sender-constrained token | ✓ | ✗
API-key / virtual-key auth, no OIDC exchange |
AI Gateway Capabilities
| Feature | KnoxCall | Portkey |
|---|---|---|
Unified API to many LLM providers One endpoint in front of multiple model vendors | ✓
AI Gateway across major providers | ✓
1,600+ models — far broader coverage |
Reliability routing (load-balance / failover / retries / semantic cache) Keep requests flowing when a provider degrades | ~
Basic failover; routing is not our headline | ✓
Mature: load-balancing, fallbacks, semantic cache |
Streaming PII redaction on prompts & completions Hold-back FSM redaction as tokens stream | ✓
FF3-1 + hold-back FSM (Pro+) | ✓
PII guardrails available |
Prompt firewall + canary leak detection Injection defence plus seeded canary tokens to catch exfiltration | ✓
Prompt firewall + canary (Pro+) | ~
Prompt-injection guardrails; canary-leak not a stated feature |
Per-agent budgets Spend ceilings scoped to an agent / key | ~
Recorded & surfaced, not hard-enforced yet | ✓
Budget & rate limits on virtual keys (granular on Enterprise) |
Observability & Cost
| Feature | KnoxCall | Portkey |
|---|---|---|
LLM logs, traces & per-request cost attribution Deep visibility into every model call | ✓
Advanced Analytics (Pro+) | ✓
Best-in-class; the reason many teams pick Portkey |
Real-time request analytics & geo Metrics dashboards and world-map request view | ✓
Analytics + geo built in | ✓
LLM-focused observability |
Alerting (Email + SMS + Slack) Notify on anomalies and thresholds | ✓
Pro+ tiers | ✓
Alerts on Production tier |
Log retention How long request logs are kept | 7–custom
7-day on Free, longer on paid tiers | 3–30d
3d Dev, 30d Production, custom Enterprise |
Secrets, Crypto & Data Protection (beyond the LLM gateway)
| Feature | KnoxCall | Portkey |
|---|---|---|
General API proxying & secrets management Not just LLMs — Stripe, Twilio, internal APIs | ✓ | ✗
LLM gateway scope |
Format-preserving tokenization (PAN / SSN / email) Shape-mimicking tokens so downstream systems stay untouched | ✓
Tokenization vaults (Pro+) | ✗ |
Encryption-as-a-Service + JWT/RSA/ECDSA/Ed25519 signing Encrypt / decrypt / rewrap with alg-confusion defence + BYOK | ✓
Crypto service (Pro+), BYOK on Enterprise | ✗ |
One-shot Ephemeral Proxy Single-use outbound request without persisting the credential | ✓ | ✗ |
Operations, Compliance & Setup
| Feature | KnoxCall | Portkey |
|---|---|---|
Managed SaaS, one bill No infrastructure to run | ✓ | ✓
Hosted, plus self-host option |
Self-hosted / open-source gateway core Run the gateway yourself | ✗
KnoxCall is proprietary SaaS, not open source | ✓
Open-source gateway, large community |
Formal compliance certifications SOC 2, HIPAA, GDPR posture | ~
SOC 2 aligned, Type II in progress; BAA available | ✓
SOC 2 Type 2, GDPR, HIPAA + BAAs on Enterprise |
Setup time Sign-up to production | Minutes | Minutes
Also fast to integrate |
Portkey and KnoxCall look adjacent because both put a gateway in front of your LLM calls, but they are solving different problems. Portkey is a control panel for production AI: it unifies 1,600+ models behind one API, routes for reliability, and gives you observability and cost attribution that are genuinely best-in-class. If your pain is “we have twelve model providers and no single view of spend, latency, or failures,” Portkey is a superb answer. KnoxCall’s pain is narrower and deeper: where does the real provider key live, and can anything in my workload read it?
Portkey’s virtual keys are a real improvement over pasting raw provider keys everywhere — they store the provider key in an encrypted vault and hand your app a virtual reference plus budget and rate controls. But the model still has to be called with the real key, and the process that makes that call is Portkey’s. Portkey guards and redacts the content flowing through; the provider secret itself continues to live in Portkey’s store and serving process. KnoxCall moves the credential boundary instead of the content boundary: the provider key is injected at the egress wire and there is no value-GET path back to it, and custodial rotation mints, verifies and deletes the provider’s own child key on a lease schedule — so the underlying vendor secret rotates, not just a store entry.
Be honest with yourself about the job. If your primary need is breadth of model coverage, world-class LLM observability, and mature reliability routing, Portkey is likely the better fit today, and we’ll say so plainly:
Ownership note for procurement: Portkey was acquired by Palo Alto Networks, with the deal closed on 29 May 2026 (Palo Alto Networks press release). That may be a plus if you already standardize on Palo Alto, or a factor to weigh if you prefer an independent vendor. We make no claim about end-of-life or roadmap changes.
Egress wire injection is not zero-residual, and we scope it tightly. The claim is specific: on the egress hot path, the real provider key never enters your workload — KnoxCall attaches it at the outbound hop and there is no endpoint that returns the value to your process. That genuinely removes the case where an RCE, a poisoned dependency, or a prompt-injected agent running printenv walks off with a long-lived vendor key.
What still lives in your workload is a KnoxCall token. It is short-lived, scoped to specific routes, audited on every call, DPoP-bindable, and revocable on demand — but until it is revoked it can route requests through the proxy. So this is a trust dependency and an extra network hop, the same tradeoff you accept with any federation or token-exchange layer, not a magic eliminator of compromise. The advantage is the difference between what leaks: a static vendor key valid for years versus a scoped, revocable proxy token. And note the boundary of the claim — it covers third-party outbound bearer keys, not your application’s own in-process encryption keys.
Two more honest limits worth stating: per-agent budgets on KnoxCall today are recorded and surfaced, not hard-enforced, and our KnoxCall SDKs are the six in our monorepo — not yet published to pip or npm. Migration of secrets into KnoxCall is import-only.
The clean way to think about it: many teams will run Portkey and KnoxCall. Keep Portkey for model breadth, routing, and observability; put KnoxCall on the credential boundary so the provider key never renders into the process at all and the underlying vendor key rotates itself. They are complementary far more than they are substitutes.
Portkey pricing is log-volume based, not per-seat: Developer (free, 10K logs, 3-day retention), Production ($49/mo, 100K logs + $9/100K extra, 30-day retention), and custom Enterprise. Source: portkey.ai/pricing (verified July 2026). Portkey was acquired by Palo Alto Networks, deal closed 29 May 2026.
In most cases, no. Portkey is a production LLM gateway focused on model coverage, reliability routing, and observability, while KnoxCall works the credential boundary so the real provider key never enters your workload and the underlying vendor child key is rotated custodially. The two are complementary far more than they are substitutes, and many teams will run both.
Yes. Keep Portkey for model breadth, routing, guardrails, and cost attribution, and put KnoxCall on the credential boundary so the provider key is injected at the egress wire instead of living in your process. Note that migration of secrets into KnoxCall is import-only.
Portkey is likely the better fit if you need one API to 1,600+ models, best-in-class LLM observability with per-request cost attribution, or mature reliability routing with load-balancing, automatic fallbacks, retries, and semantic caching. It also wins if you need formal compliance certifications now, since its Enterprise tier carries SOC 2 Type 2, GDPR, and HIPAA with custom BAAs and VPC hosting, or if you want to self-host an open-source gateway. KnoxCall is SOC 2 aligned with Type II in progress, and it is proprietary managed SaaS rather than open source.
KnoxCall is tiered managed SaaS: a Free Forever plan, Starter at $19/mo, Pro at $99/mo, and a custom Enterprise tier. Portkey's pricing is log-volume based rather than per-seat: a free Developer tier with 10,000 recorded logs per month, Production at $49/mo for 100,000 logs plus $9 per additional 100K requests, and a custom Enterprise tier. Portkey also offers a free open-source gateway core that you self-host and operate yourself, which KnoxCall does not.
Run KnoxCall on the credential boundary so the real vendor key never enters your workload and rotates itself — while Portkey keeps doing routing, guardrails, and cost attribution. Complementary, not rip-and-replace.