Portkey

• Palo Alto Networks could narrow future releases of the gateway toward Prisma AIRS or change the license on them, which would leave the model-neutral routing that drew developers only in the versions already published under the MIT license. The repository points the other way for now, carrying a pre-release notice that the core enterprise gateway is merging into open source with the 2.0 release.
• The public Portkey-AI/gateway repository has recorded no push since May 25, 2026, four days before the acquisition closed, so a continued freeze could move its developer community to a still-independent alternative such as LiteLLM.
• A buyer standardizing on several model vendors may decline a gateway owned by one security vendor, which would cost the product the neutral position that made enterprises route competing providers through it.
• The commercial record rests on testimonials the vendor publishes and usage figures that measure its own platform, so a disclosed post-acquisition customer count could land well below what the purchase price implies.
• The founding team now works inside a large platform vendor, so the shipping cadence that produced the gateway is now set by that vendor's priorities rather than by the founders'.

Portkey sells a control plane for the teams putting generative AI into production. It frames the buyer as the AI teams that need everything to go to production, and describes its job as connecting, managing, and securing AI interactions across more than 3,000 models with centralized control, real-time monitoring, and governance. The product sits between the application and the model providers, so failures and spend become visible in one place.

An independent analyst corroborates the problem at enterprise scale, and the acquirer's framing does not. Futurum analyst Fernando Montenegro wrote on May 7, 2026 that the gateway addresses a real and growing control gap as agentic AI workloads move from pilot to production. Palo Alto Networks, announcing its intent to buy, described a centralized control plane for managing and protecting autonomous AI agents, but it now owns the problem, so its account is a buyer's conviction rather than outside evidence.

The asset under management is the AI request itself. Every prompt an application sends and every response it gets back crosses the gateway, which is the point where untrusted input meets privileged model access, and that is the boundary the agent-security wave has made urgent. [s1, s2, s16, s7]

The core of Portkey is an open-source gateway with commercial layers above it. The public Portkey-AI/gateway repository carries 3,457 commits, 12,842 stars, and an MIT license, and describes routing to 1,600-plus models with 50-plus guardrails behind one API. The New Stack reported on March 31, 2026 that the company had launched its newly unified Portkey Gateway service and made it fully open source, while the repository still describes an Enterprise Version with organization, governance, security, and deployment capabilities the free build does not carry. The same README carries a pre-release notice that the core enterprise gateway is merging into open source with the 2.0 release, so the split the Enterprise Version describes is one the project says it is closing.

That open code also carries a public vulnerability record. A request-forgery weakness in the gateway's custom-host header handling is tracked as CVE-2025-66405, which Portkey fixed in version 1.14.0. The National Vulnerability Database scores it 9.8 and marks it analyzed, the GitHub advisory service that assigned it scores the same flaw 6.9, and a CISA coordinator recorded exploitation as none and automatable as no.

Above the gateway sit guardrails, observability, and governance. Portkey says it runs more than 60 guardrails on the open-source gateway to filter, fix, or route each request as it happens, aimed at injections and data leaks, alongside conditional routing, fallbacks, retries, and load balancing that address the reliability problems it names.

What ships under the acquirer's name is the same layer with more around it. Palo Alto Networks made the Prisma AIRS AI Gateway generally available on July 16, 2026, describing it as sitting inline between every AI interaction, model provider, and agentic interaction, with a single enforcement point for operational and security controls. [s5, s19, s14, s10, s3, s2, s21]

Portkey competed in AI gateways against open-source rivals and the platforms bundling the category. The New Stack named Portkey, LiteLLM, Kong AI Gateway, and Cloudflare AI Gateway as products that each solved the same model-fragmentation problem in slightly different ways, and LiteLLM's own documentation pitches calling more than 100 models in the OpenAI format through a central gateway service.

Portkey's entry point was a free gateway with broad model coverage sitting beneath a paid control plane. Anyone can run the open build locally, and the commercial layer carries the observability and policy an enterprise pays for. The reviewed record documents that structure and does not document how many of those free users became paying buyers.

The category question the market answered was whether a standalone gateway stays independent. Palo Alto Networks closed its purchase on May 29, 2026 and made the gateway the core of Prisma AIRS. For Portkey, that resolved with the gateway arriving inside a larger security platform rather than holding the field alone. [s15, s23, s6, s5]

The named accounts Portkey names are testimonials on its own pages. The homepage carries quotes attributed by name and title to people at Qoala, whose chief technology officer cites 30 million policies a month across more than 25 generative-AI use cases, and at QA.tech, Figg, Ario, Internet2, Perficient, RVO Health, and Snorkel AI, alongside an unnamed Fortune 500 pharmaceutical AI leader. The pricing page adds Haptik's chief technology officer, who says the company runs Portkey in staging and production. These are quotes the vendor publishes rather than contracts it discloses.

One account appears on both sides of the record. The independent client list published in a March 2024 Inc42 profile names Turing, Haptik, Postman, Quizizz, Springworks, and Merlin as clients onboarded since incorporation, and Haptik is the account the vendor and that outside outlet both name.

Usage scale is the strongest number in the record and it came from the company. The New Stack reported at the March 2026 open-source launch that the platform was already processing trillions of tokens and more than 120 million AI requests every day, managing $180 million in annualized AI spend across some 24,000 organizations. Those numbers describe Portkey's own platform, and no source in the reviewed record measures them independently.

The distribution the product reaches now belongs to the acquirer. Palo Alto Networks made the gateway generally available inside Prisma AIRS on July 16, 2026 and sells it through its own enterprise motion, which is reach the buyer supplies rather than traction the gateway earned. [s1, s4, s18, s14, s21]

Portkey was built by two founders whose record is in AI product and infrastructure. Rohit Agarwal led the product and AI vertical at Freshworks from 2015 to 2020, then scaled the AI writing assistant Peppertype.ai at Pepper Content, and in April 2023 he co-founded Portkey with Pepper Content's former director of engineering Ayush Garg to build a control panel for AI applications.

The pair now have a priced exit on that record. Palo Alto Networks' Form 10-Q states the purchase at $140 million in cash and replacement awards, so an outcome that press coverage described as undisclosed is documented in a regulatory filing, and it sits in the same field as the earlier build.

The investor signal ran alongside it. Lightspeed led the 2023 seed, with its partner Dev Khare saying he had known Rohit and Ayush for a decade through multiple efforts from zero to one, and returned for the 2026 Series A that Elevation Capital led. [s18, s20, s17, s8, s9]

Portkey lists its attestations on a dedicated trust surface. The trust portal at trust.portkey.ai names HIPAA, SOC 2, GDPR, and ISO 27001:2022, each with one document, and gates the documents themselves behind a request for access. The pricing page's Enterprise tier names SOC 2 Type 2, GDPR, and HIPAA compliance alongside custom business-associate agreements and data isolation. The reviewed record shows no federal authorization or regulatory mandate attached to the product, so the assurance on offer is the commercial-audit kind.

Deployment control is the part of the story the architecture carries. The gateway can run self-hosted from the MIT-licensed build or in a private cloud with virtual-private-cloud hosting and custom retention periods for logs and metrics on the Enterprise tier, so an enterprise can run the gateway itself rather than relying on an attestation alone.

What the deal changes is who sets the public project's cadence. The GitHub API records the last push to Portkey-AI/gateway on May 25, 2026, four days before the acquisition closed, and the repository is not archived, so the license and the model-neutral breadth that earned developer trust now sit with a single security vendor. [s22, s4, s5, s19, s25]

Add analyzed competitors to compare them side by side with Portkey.

Portkey targets the AI teams taking generative AI into production. The vendor frames the buyer as teams needing one place for the gateway, observability, guardrails, governance, and prompt management, so the segment is defined by the production-AI problem rather than by a regulated vertical. That places the buyer where untrusted input meets privileged model access.

An independent analyst corroborates the segment at enterprise scale. Futurum's Fernando Montenegro wrote on May 7, 2026 that the gateway addresses a real and growing control gap as agentic AI workloads move from pilot to production, and he treats whether organizations running Prisma AIRS activate the capability at scale as still open. Palo Alto Networks describes the same buyer, but it now owns the product, so its account carries a purchaser's conviction rather than outside evidence.

The adoption set the record shows spans startups through large enterprises, and it is the vendor's own account. The homepage carries testimonials from Qoala, whose chief technology officer cites 30 million policies a month across more than 25 generative-AI use cases, alongside QA.tech, Figg, Ario, Internet2, Perficient, RVO Health, Snorkel AI, and an unnamed Fortune 500 pharmaceutical AI leader. Two of those speakers describe running the gateway themselves, Perficient's vice president of AI data platforms strategy in a hybrid deployment and, on the pricing page, Haptik's chief technology officer in staging and production. These are quotes the vendor publishes rather than disclosed contracts, so they indicate reach across the segment without stating what any named account pays.

The gateway routes and governs model traffic across many providers behind one interface. The feature page describes connecting, managing, and securing AI interactions across more than 3,000 models with centralized control, real-time monitoring, and governance, and adds conditional routing, fallbacks between models on failure, automatic retries, and load balancing aimed at the reliability problems production teams hit.

The core is code a buyer can read, run, and fork. The public Portkey-AI/gateway repository carries 3,457 commits, 12,842 stars, and an MIT license, and describes routing to 1,600-plus models with 50-plus guardrails behind one API. The New Stack reported on March 31, 2026 that the company had launched its newly unified Portkey Gateway service and made it fully open source, while the repository still describes an Enterprise Version carrying organization, governance, security, and deployment capabilities the free build does not. The same README carries a pre-release notice that the core enterprise gateway is merging into open source with the 2.0 release, so the project says it is closing that split.

Above that base sit the checks the vendor sells. Portkey says it runs more than 60 guardrails on the open-source gateway to filter, fix, or route each request as it happens, aimed at injections and data leaks.

The advantage is method and placement rather than a retained data asset. The routing, guardrail, and observability work is real engineering, and the controls are configured policies over inspectable, forkable code, so another funded team could reproduce the capability by writing similar software. Palo Alto Networks separately characterizes the architecture as able to process trillions of tokens, which is the purchaser's framing and not independent measurement of the gateway.

Portkey's route to market starts with free software and moves upward. The open-source gateway and a free Developer tier pull engineers in before any sales conversation, and the site then names enterprise customers as social proof for the paid platform. That is the motion the line controls by shipping software and earning references, though the fetched pages document the motion rather than any conversion rate from free users to paying ones.

The customer proof the line carries is mostly vendor-published, and the independent proof is dated. The homepage names Qoala running 30 million policies a month across more than 25 generative-AI use cases, alongside QA.tech, Figg, Ario, Internet2, Perficient, RVO Health, Snorkel AI, and an unnamed Fortune 500 pharmaceutical AI leader, and the pricing page adds Haptik. A March 2024 Inc42 profile named Turing, Haptik, Postman, Quizizz, Springworks, and Merlin among as many as 15 clients onboarded since incorporation, which makes Haptik the one account the vendor and an outside outlet both name.

The usage numbers are large and they measure the company's own platform. The New Stack reported at the March 2026 open-source launch that the platform was already processing trillions of tokens and more than 120 million AI requests every day, managing $180 million in annualized AI spend across some 24,000 organizations. Those numbers describe Portkey's own platform, and no source in the reviewed record measures them independently.

Distribution now belongs to the acquirer, and the paid roster is still thin in public. Palo Alto Networks made the gateway generally available inside Prisma AIRS on July 16, 2026 and sells it through its own enterprise motion, which is reach the buyer supplies. A current count of paying enterprises, contract terms, and post-acquisition retention are absent from the fetched record.

Portkey publishes a rate card a buyer can read without a call. The pricing page lists a free Developer tier for prototyping with 10 thousand recorded logs a month, a Production plan at 49 dollars a month with 100 thousand recorded logs and 9 dollars per additional 100 thousand requests, and a custom Enterprise tier carrying more than 10 million recorded logs a month.

The metered unit is the request, which matches how a production-AI team already counts its own traffic. Charging by recorded logs and request volume rather than per seat lets usage grow without a renegotiation, which fits the motion the free gateway starts.

The published price stops where procurement begins. The Production plan states it is not recommended for organizations requiring custom security controls or data residency guarantees, which routes those buyers to the Enterprise tier holding the compliance certificates, private cloud deployment, custom retention, and data isolation. That is the two-track shape a product sold into gated accounts needs.

Portkey delivers as a hosted service with a build the customer can run itself. The pricing page offers the hosted Production plan and an open-source option a buyer deploys locally, and the Enterprise tier adds private cloud deployment and virtual-private-cloud hosting, so the operational load the customer carries scales with how much control it wants over where requests travel.

The operational core is inline request handling rather than after-the-fact analysis. The gateway sits in the path of every model call, applying routing, fallbacks, retries, load balancing, and guardrails as each request happens, which is what lets it catch failures and policy violations in real time and also makes it a dependency in the production path. Palo Alto Networks describes the same placement for the shipped product, inline between every AI interaction, model provider, and agentic interaction.

That placement raises the question a security review opens with. A gateway every model request passes through becomes a component whose latency or outage stops production AI, and the pricing page claims enterprise-grade reliability backed by robust service-level agreements while the fetched pages carry neither the agreement terms nor any independent latency measurement.

Portkey lists its attestations on a dedicated surface. The trust portal at trust.portkey.ai names HIPAA, SOC 2, GDPR, and ISO 27001:2022, each with one document, and gates the documents themselves behind a request for access. The pricing page's Enterprise tier names SOC 2 Type 2, GDPR, and HIPAA compliance alongside custom business-associate agreements and data isolation. The reviewed record shows no federal authorization or regulatory mandate attached to the product.

The architecture carries the part an attestation cannot. The gateway can run self-hosted from the MIT-licensed build or inside a private cloud, with data isolation and custom retention periods for logs and metrics on the Enterprise tier, so a customer chooses where the gateway itself runs. The fetched pages present these as deployment options rather than as a stated residency guarantee, and they do not describe where a request travels once the gateway forwards it to a model provider.

The certificates and the self-hostable build are what a buyer can verify without reference to the parent. Palo Alto Networks brings a large security vendor's procurement and support posture behind the line, which answers the vendor-viability worry a single startup gateway raises, and that posture is the acquirer's to lend rather than something the gateway built. What it changes is who sets the public project's cadence, since the GitHub API records the last push to the repository on May 25, 2026, four days before the deal closed.

Portkey is built to be the layer other applications route their model traffic through. The vendor frames the gateway, observability, guardrails, governance, and prompt management as one platform sitting between applications and more than 3,000 models, so the systems consuming model output depend on the gateway to reach those models at all.

The open build widens what surrounds it. The public repository routes across 1,600-plus models and runs guardrails developers can extend, and it carries 1,270 forks alongside its 12,842 stars, so engineers build against Portkey rather than only consuming it. The open-source launch that drew them came in March 2026, before the acquisition closed.

Inward, the platform is now a component of a larger suite. Palo Alto Networks made the Prisma AIRS AI Gateway generally available on July 16, 2026, describing it as a unified gateway for model, tool, and agent-to-agent traffic with a single enforcement point, which widens distribution while tying the roadmap to one security vendor's priorities. The public repository has recorded no push since May 25, 2026 and is not archived.

Portkey was built by founders whose background is AI product and infrastructure rather than security. Rohit Agarwal led the product and AI vertical at Freshworks from 2015 to 2020 and then scaled the AI writing assistant Peppertype.ai at Pepper Content, and in April 2023 he co-founded Portkey with Pepper Content's former director of engineering Ayush Garg to build a control panel for AI applications.

The documented expertise sits in AI operations, which is the problem the gateway solves. Agarwal names the LLM hallucinations and inefficient AI operations that peeved him at Freshworks, and the small internal tools he and his team built there and at Pepper Content that solved only part of the problem. Palo Alto Networks now positions the same gateway inside its security platform.

The pair now carry a priced outcome. Palo Alto Networks' Form 10-Q states the purchase at $140 million in cash and replacement awards, so a result that press coverage described as undisclosed sits in a regulatory filing, next to a named earlier build in the same field. Lightspeed led the 2023 seed and returned for the 2026 Series A that Elevation Capital led.

After the deal the line sits inside a public company. Palo Alto Networks' quarterly filing records the purchase and the integration into Prisma AIRS, so the vendor behind the gateway is now a large one, which is the acquirer's standing rather than capacity the founding team held.