Compare TrueFoundry vs Portkey
Palo Alto Networks announced its acquisition of Portkey on April 30, 2026. TrueFoundry is an independent, Kubernetes-native AI Gateway, MCP Gateway and Agent Gateway that runs entirely inside a team's VPC.
On guardrails such as PII redaction, TrueFoundry has built-in, in-process PII/PHI and secrets detection with no external dependencies, plus four guardrail hooks (LLM input, LLM output, MCP pre-tool, MCP post-tool). Portkey's native MCP guardrails are still in early access, and its guardrail configuration stays in Portkey's cloud.
When TrueFoundry Makes Sense
TrueFoundry is the only independent, full-stack AI gateway built for integration, observability, cost management, and agentic AI. TrueFoundry combines a high-performance AI Gateway (~3ms latency), MCP Gateway, Agent Gateway, and full model deployment into one Kubernetes-native system that runs entirely inside a team's VPC.
Teams still comparing can test TrueFoundry on their own traffic, with no credit card required.
Key Competitive Differentiators
TrueFoundry and Portkey compared across twelve capabilities.
| Capability | TrueFoundry | Portkey |
|---|---|---|
| Vendor Independence | TrueFoundry's roadmap is driven entirely by enterprise AI infrastructure needs, not by a parent company's security platform strategy. A team's deployment, support, and product direction stay stable. | Palo Alto Networks announced its acquisition of Portkey on April 30, 2026. |
| Gateway Architecture & Performance | Enterprise-grade, Kubernetes-native gateway with ~3ms latency at 250 RPS per pod, scaling linearly to tens of thousands of RPS. Stateless hot path with in-memory auth, rate limiting, and guardrail enforcement. | Hybrid VPC data plane keeps LLM payloads in-network; control plane and dashboard remain SaaS-hosted. |
| Routing & Load Balancing | Native latency-based & weight-based routing using inter-token latency / TPOT, adaptive priority with SLA cutoff, weight-based routing, typed YAML policies, and OTEL export. Routing is configurable at team, model, and application level with four-hook guardrails on every path. | Routing scoped at workspace level. Best for SaaS teams wanting composable routing without K8s expertise. |
| Deployment Options | Fully Kubernetes-native deployment in the customer's VPC (cloud or on-prem). The entire hot path — auth, rate limits, guardrails — runs in-cluster with no external dependencies. | Hybrid VPC data plane keeps prompt payloads in the customer network. However, control-plane sovereignty is not achievable in hybrid mode: the dashboard, guardrail config UI, and analytics aggregation remain in Portkey's cloud. |
| Data Residency & Sovereignty | Complete data sovereignty: in-memory request handling, four-hook guardrails with composable per-team masking, built-in in-process PII/PHI and secrets detection, OTEL trace export to customer-controlled backends, and customer-managed storage options. | Hybrid VPC mode genuinely keeps prompt payloads in the customer network. The defined architectural tradeoff: control-plane sovereignty is not achievable — the dashboard, guardrail config, and analytics aggregation remain Portkey-hosted by design. |
| MCP Gateway | Purpose-built enterprise MCP governance: dedicated pre/post-tool guardrail hooks, support for public/self-hosted/custom MCPs, Virtual MCP Servers. | Native MCP guardrails are still in early access — custom validation is via adjacent webhook path, not a first-class MCP policy engine. |
| Agent Gateway | Four-hook guardrail system (LLM input, LLM output, MCP pre-tool, MCP post-tool) is especially complete for tool-calling agents. Split-plane design keeps the gateway governing traffic while async services handle long-running loops. | No native async execution substrate, requiring application-side orchestration for long-running agent loops. |
| Guardrails | Subject-scoped rules, metadata-scoped rate limits, MCP per-invocation hooks, built-in PII/PHI and secrets detection — zero external dependencies. Covers HIPAA, GDPR, GovCloud, and air-gap. | Native MCP guardrails are not yet GA. For teams with compliance requirements, the control plane — including guardrail configuration — remains in Portkey's cloud, meaning policy changes flow through vendor-hosted infrastructure regardless of deployment mode. |
| Observability | Full-stack observability: OTEL export, Prometheus/Grafana integration, and built-in Metrics Dashboard. | Built-in request logging, token usage and cost tracking dashboard (real-time). Limited visibility into underlying infra (since it doesn't host models). |
| Cost Control | Hot-path budget enforcement (not reactive). Cost attribution by team/user/model/app across external APIs and self-hosted models. 35–50% TCO reduction documented. | Workspace-level budget governance, custom model pricing, and per-provider cost tracking. Budgets are reactive (post-spend). No self-hosted model cost management. |
| Ecosystem Integration | Broad integration: works within a team's CI/CD, GitOps pipelines; connects to Kafka/SQS for async pipelines. Plays nicely with cloud services (AWS, GCP) but remains cloud-agnostic. Open APIs to integrate custom tools. | Developer-centric integrations: ready connectors for LangChain, LlamaIndex, Flowise, etc., to plug into LLM apps. Less integration for non-LLM workflows (e.g., ETL or CI/CD). |
| Support | 24×7 via Slack and on-call engineers, dedicated AM. G2 rating 9.9/10. SOC2 and HIPAA compliant. | Community-driven support (Discord/GitHub for OSS). Enterprise plan offers support SLAs, but overall smaller support setup (startup scale). |
How data residency differs
TrueFoundry runs the entire hot path — auth, rate limits, guardrails, traces — inside a team's Kubernetes cluster with no external dependencies. Portkey's hybrid VPC mode genuinely keeps inference payloads in-network, but the dashboard, guardrail config UI, and analytics aggregation remain in Portkey's cloud by design. This is a well-documented architectural tradeoff, not a hidden gap — but it is a hard constraint for teams requiring full control-plane residency.
Key Evaluation Questions
Five questions buyers ask when evaluating TrueFoundry against Portkey, with how TrueFoundry addresses each and the Portkey considerations.
| Question | How TrueFoundry Fixes It | Portkey considerations |
|---|---|---|
| "How do we get full data sovereignty, no payload or control-plane egress." | Built-in PII/PHI and secrets detection requires no external services. OTEL traces export to a team's own backends. Customer-managed storage options available. | Hybrid VPC mode keeps inference payloads in-network, but the dashboard, guardrail config UI, and analytics aggregation remain in Portkey's cloud by design. |
| "Can we optimize LLM costs while also controlling self-hosted model spend?" | TrueFoundry provides Public/Private Cost pricing for internal chargebacks across both external APIs and self-hosted models. Documented 35–50% TCO reduction through K8s workload optimization and spot/GPU scheduling. | Portkey offers workspace-level budget controls and model catalog pricing with per-provider cost tracking. Budgets are reactive (post-spend), and unsupported models require manual pricing configuration. No self-hosted model deployment or cost optimization capabilities. |
| "We need MCP governance for production agentic workloads." | TrueFoundry MCP Gateway provides dedicated pre/post-tool guardrail hooks, Virtual MCP Servers, Cedar-based policy engine, OAuth 2.0 and inbound OAuth for the gateway, Secret Groups for credential isolation, and full MCP-specific observability — all inside a team's VPC. | Portkey's MCP Gateway still does not have native MCP guardrails, so custom tool-call validation relies on an adjacent webhook path rather than a first-class policy engine. |
| "Do we need observability across LLM calls and our own deployed models?" | TrueFoundry offers end-to-end observability – not only do teams get request metrics, but also container logs, live monitoring, and alerts down to the pod level. Developers can debug failures through a UI, inspect logs in real-time, and even profile models. This holistic view speeds up troubleshooting significantly. | Portkey gives good LLM-level observability (token counts, latencies, errors) via its dashboard. But it won't trace issues inside a custom model container – that's outside its scope. Debugging infrastructure failures or performance bottlenecks in a team's own model server is manual. |
| "Will we outgrow the platform's capabilities?" | TrueFoundry is built for this evolution. It manages both external API routing and internal self-hosted model deployment from one interface, so migrating from OpenAI to a private Llama model doesn't require a platform change or application rewrites. Training, fine-tuning, serving, and gateway are all unified. And as an independent company, TrueFoundry's roadmap is entirely focused on enterprise AI infrastructure, so the platform grows in the direction each team needs. | Palo Alto Networks announced its acquisition of Portkey on April 30, 2026. Teams building a long-term AI infrastructure strategy should ask whether a security-platform acquisition will continue to prioritize developer-centric AI gateway capabilities, or whether those needs will become secondary to Palo Alto's broader enterprise security agenda. |
Which solution provides more advanced debugging tools
TrueFoundry connects LLM request traces with infrastructure metrics in a single UI — GPU memory, pod health, container logs. When something breaks, a team can see whether it's a model issue (bad prompt, token overflow) or an infrastructure problem (OOM error, pod failure) without leaving the platform. Portkey delivers excellent LLM-level observability and now includes Langfuse/LangSmith integrations and enterprise export paths — but since it doesn't host models, infrastructure-level failures in a team's model server are outside its scope entirely.
Evaluating TrueFoundry on real requests
A team can add its provider keys and route real requests through TrueFoundry's routing, budgets, and guardrails, and know within the hour. The free tier covers 3 users, with SOC 2 from day one and no credit card required.
How TrueFoundry acts as a Painkiller
Five pain points, the benefit TrueFoundry brings to each, and the customer impact.
| Key Painpoints | Benefits of using TrueFoundry | Customer Impact |
|---|---|---|
| Fragmented AI Infrastructure | One platform covers model serving, LLM / MCP / Agent Gateway, prompt management, guardrails, observability, and cost control. No context-switching between tools, no duplicate configuration across systems. | Engineering teams ship AI features faster with fewer handoffs. Platform teams manage one system instead of five. The entire AI stack is visible, debuggable, and governed from a single control plane. |
| Incomplete Data Sovereignty | Full-stack residency: in-memory hot path, built-in PII/PHI and secrets detection, composable per-team masking policies, OTEL export to customer-controlled backends, and customer-managed storage. Nothing leaves a team's environment. | Regulated industries can deploy AI in production with confidence. Security and compliance teams get audit-ready infrastructure out of the box. No renegotiating data processing agreements every time a team adds a new model or tool. |
| Reactive Cost Governance | Cost attribution runs across teams, users, models, and applications — for both external API calls and self-hosted model fleets. Public/Private Cost pricing enables accurate internal chargebacks. | Finance and platform teams gain a real-time, unified view of AI spend across every team and model. Budget overruns are prevented rather than discovered. Internal chargeback becomes straightforward, making AI costs visible and accountable at the team level. |
| Limited MCP & Agent Governance | Deep observability built-in: real-time logs, detailed error traces, and performance metrics for every request. TrueFoundry's UI and alerts enable quick root-cause analysis (whether it's a bad prompt, a slow model, or infrastructure glitch), minimizing downtime and improving reliability. | Blind spots in production – teams struggle to pinpoint issues with prompts or model performance. Minimal logging from external APIs; homegrown model servers lack unified monitoring, leading to prolonged downtimes. |
| Vendor and Platform Lock-In | Cloud-agnostic and Kubernetes-native. Deploy on any cloud or on-prem. Supports any model, library, or framework. As an independent, founder-led company, TrueFoundry's roadmap is driven entirely by enterprise AI infrastructure needs — with no parent company's security agenda to accommodate. | Teams retain full optionality as the AI landscape evolves. Switching models, adding new providers, or moving clouds never requires a platform migration. Leadership can make infrastructure decisions based on business needs — not vendor constraints. |
Common Pitfalls to Avoid
Six pitfalls avoided by using a cloud agnostic platform such as TrueFoundry over Portkey.
- Underestimating what data sovereignty actually requires. Keeping inference payloads in-network is a good start. But if a team's guardrail configuration, analytics, and dashboard live in a vendor's cloud, the team's control plane is still exposed. For regulated industries, that distinction matters at audit time.
- Underestimating MCP governance maturity requirements. Central server onboarding and OAuth support are table stakes. Production agentic workloads require per-tool guardrail hooks, policy enforcement, and credential isolation. Before standardizing on a platform, verify which of those capabilities are GA and which are still on the roadmap.
- Overlooking vendor independence. When a gateway tool gets acquired by a security platform, the roadmap shifts toward the acquirer's priorities, not the team's. Choose a platform whose incentives stay aligned with the team's own.
- Mistaking reactive budgets for cost control. Getting alerted after a team has already overspent isn't governance, it's accounting. Real cost control means enforcing budgets on the hot path, with attribution across every team, model, and application before the bill arrives.
- Building agent infrastructure on a per-call proxy. Retries, fallbacks, and circuit breakers handle individual calls well. But long-running agents need a native async execution substrate. Without one, a team ends up building and maintaining that orchestration layer on top of the gateway indefinitely.
- Conflating developer velocity with enterprise readiness. A near-zero-config dashboard and UI-driven guardrails are great for POCs. Physical compute isolation, two-layer RBAC, Kubernetes namespace boundaries, and GitOps-native pipelines are what get teams to production and keep them there.
Real Outcomes at TrueFoundry
Results delivered by TrueFoundry for five customers. Customers shown on truefoundry.com include Automation Anywhere, Siemens Healthineers, Resmed and Innovaccer.
- Automation Anywhere deploys a multi-region LLM gateway deployment and has set up RBAC for model and MCP access through the gateway.
- Siemens Healthineers controls model access and does chargeback to teams through cost accounting.
- Resmed is exploring and using TrueFoundry for multiple use cases.
- Innovaccer routes all AI inference calls across experimentation and production, processing over 1 billion tokens monthly across ~10 applications.
- One customer manages and routes inference across multiple models, including self-hosted ones, handling requests with production-grade reliability.
Related reading
Independent company. AI infrastructure only.
See the full platform on the free tier (gateway, MCP, prompts, deployment). No credit card required. SOC 2. G2 9.9/10.