Almost every gateway in this category calls itself open source. Almost none of them means the same thing by it.
Some are open source with a commercial cloud on top and nothing withheld. Some are open source except for a directory that happens to contain the governance features. Some are open source but put high availability on the paid tier. All three are legitimate business models. Only one of them is what a team means when they say "we need something we can self-host".
So this post asks one question of each gateway and answers it from primary sources: what stops working if you self-host and never pay?
Everything below was checked against each project's LICENSE file, repository and pricing page during August 2026. Star counts and prices move; re-check before you commit.
LiteLLM
Stars: 57,637. Licence: MIT, except enterprise/.
The default answer in this category, and it earns that. Widest provider coverage anywhere, a Rust core with a Python SDK, and commits landing daily.
The licence is split, and LiteLLM is upfront about it. The root LICENSE states that everything under the enterprise/ directory is governed by a separate licence and everything else is MIT. That directory exists, which is why GitHub reports the repository licence as NOASSERTION rather than MIT.
What is on the paid side: virtual keys, budgets, RPM and TPM limits, spend tracking by key, user, team and org, SSO with SCIM, and OIDC or JWT auth. In other words, governance.
That matters because governance is not what you need on day one, it is what you need in month six when a second team starts using your gateway. Enterprise pricing is not published; the site offers a 30 day trial key and a sales conversation.
Pick it when: you need breadth of provider support, or the MIT core covers you and always will.
Bifrost
Stars: 7,675. Licence: Apache-2.0, no enterprise directory.
Maxim AI's gateway, and a genuinely well-engineered one. A Go binary built for low overhead, claiming under 100 microseconds at 5,000 requests per second and 50x faster than LiteLLM. Those are their numbers and we have not benchmarked them, but the engineering focus is real and the deployment story is simple: Docker, Kubernetes, or a single binary.
The licence is clean. No carve-out directory, which is a better structure than LiteLLM's.
What is on the paid side: guardrails, Cluster Mode for high availability, Adaptive Load Balancing with real-time health monitoring, RBAC, audit logs, log exports, SSO via SAML and OIDC, and vault integrations.
Read that list carefully. Cluster Mode is high availability. Putting HA and guardrails on the Enterprise tier means the free build is closer to a development gateway than a production one. That is a legitimate model and the licence stays honest, but it is worth pricing before adoption rather than after.
Pick it when: raw proxy overhead is your binding constraint, or the OSS tier genuinely covers you.
Kong AI Gateway
Stars: 44,059. Licence: Apache-2.0 core.
Kong Gateway itself is properly Apache-2.0 and has years of production hardening behind it. If you already run Kong for your APIs, adding AI traffic to a gateway your team already operates is a strong argument that beats most feature comparisons.
The AI capabilities arrive as plugins, and the split is where it gets interesting. Six are free: ai-proxy, ai-prompt-guard, ai-request-transformer, ai-response-transformer, ai-prompt-template and ai-prompt-decorator.
Seventeen require an AI Gateway Enterprise licence, each carrying an explicit "AI License Required" badge on its own documentation page: ai-proxy-advanced, ai-semantic-cache, ai-semantic-prompt-guard, ai-semantic-response-guard, ai-rate-limiting-advanced, ai-sanitizer, ai-rag-injector, ai-prompt-compressor, ai-mcp-proxy, ai-mcp-oauth2, ai-llm-as-judge, ai-lakera-guard, ai-aws-guardrails, ai-azure-content-safety, ai-gcp-model-armor, ai-custom-guardrail and ai-a2a-proxy.
So semantic caching, advanced rate limiting and every guardrail are paid.
There is also a per-model fee: Konnect Plus includes five unique proxied LLMs, then charges $100 per month for each additional model. Multi-provider routing exists so you can add models freely, so metering the model count prices the thing the gateway is for.
One more thing worth knowing: Kong's own July 2026 post states that Kong Gateway 3.18 makes AI plugins opt-in rather than bundled, and that AI Gateway 2.0 moves to its own runtime and release cadence. The packaging is explicitly changing.
Pick it when: Kong is already in your path, or you need a general-purpose API gateway too.
Helicone
Stars: 6,116. Licence: Apache-2.0, no enterprise directory. In maintenance mode since March 2026.
One of the cleanest licence stories here. The maintenance status is easy to get wrong in either direction, so here it is precisely.
Mintlify acquired Helicone in posts both companies published on 3 March 2026. Helicone's says "Helicone has been acquired by Mintlify, and our team will be joining them in San Francisco", that "Helicone's services will remain live for the foreseeable future in maintenance mode", and that "This means security updates, new models, bug & performance fixes all keep shipping." Mintlify's post says the same thing. So maintenance mode is the company's own word for it, and their definition of it explicitly is not abandonment.
The commit history matches both halves of that. On main: 50 commits in November 2025, 76 in December, 39 in January 2026, 37 in February. Then 11 in March, when the acquisition was announced, and 2, 6, 1, 1 and 3 in the five months since. Commits still land, and the cadence collapsed exactly when the deal did.
One more thing worth knowing before you evaluate it: the repository is Apache-2.0, is not archived, and its README mentions none of the above. A reader landing there sees a normal, active project.
It is both a gateway and an observability platform, and integration is about as low-friction as this category gets: a base URL change and an auth header.
What is tier-gated: supported on-prem deployment sits under Enterprise, and SOC 2 and HIPAA compliance sit on Team at $799 per month. The Apache-2.0 code is available regardless, so the distinction is between running the software and having a vendor relationship for it.
Pricing: Hobby free with 10,000 requests and 1 GB storage, Pro $79 per month with unlimited seats, Team $799 per month, Enterprise contact-us.
Pick it when: observability plus routing is the whole problem, the hosted tiers suit you, and a product whose vendor has committed to fixes and new models rather than new features is an acceptable place to build on.
OpenRouter
Not open source. Not self-hostable.
It belongs on this list because it comes up in every one of these searches, and because it is genuinely excellent at its actual job, which is not the one this post is about.
OpenRouter aggregates the widest model catalogue available behind one key, and passes inference through without markup, which is a fairer commercial model than most. It charges on credit purchases instead: 5.5% with a $0.80 minimum via Stripe, or 5% via crypto. Bringing your own provider keys is free up to $25,000 of monthly list price on pay-as-you-go and $200,000 on enterprise, then 5% of standard OpenRouter pricing above that.
But its documentation describes no self-hosting or open source option, so prompts and responses transit their infrastructure.
Pick it when: you are evaluating models rather than running constrained production traffic. For breadth, nothing self-hosted competes.
Cloudflare AI Gateway
Not open source. Not self-hostable.
Same caveat, same reason for inclusion. It is a Cloudflare platform service: analytics and logging, caching served from Cloudflare's edge instead of the model provider, rate limiting, and retry with model fallback.
Two things make it genuinely hard to argue with. It is available on all Cloudflare plans, including free, and there is nothing to operate. Cloudflare's edge is better run than most self-hosted infrastructure will ever be.
Pick it when: you already run on Cloudflare and have no data residency constraint. If both are true, standing up your own gateway instead is difficult to justify.
Everstack
Licence: Apache-2.0 across the repository, including the enterprise modules.
Full disclosure, we build this. So judge the claim rather than the source: the specific, checkable statement is that there is no separately licensed directory, and that semantic caching, per-tenant rate limits, guardrails, virtual keys, SSO, audit logs and clustering are all in the build you can run for free.
Where Everstack differs structurally is scope. The same control plane runs a stateful agent runtime, isolated Docker, Firecracker and Kubernetes sandboxes, vector memory, workflows and evaluations, so a trace follows the work past the request boundary rather than stopping at it.
Do not pick it if routing is your whole problem. A gateway is the right size of tool for a gateway-sized problem, and LiteLLM or Bifrost will get you there with a larger community behind them. The case for a platform needs several of these problems at once, plus a deployment constraint.
How to actually choose
Three questions, in order.
1. What stops working if you self-host and never pay?
| Gateway | Free build gives you | | ---------- | --------------------------------------------------------------- | | Bifrost | Routing, but not guardrails, HA, RBAC or audit logs | | Kong | Six AI plugins, but not caching, advanced limits or guardrails | | LiteLLM | Routing and observability, but not virtual keys, budgets or SSO | | Helicone | Everything, though supported on-prem is Enterprise | | Everstack | Everything | | OpenRouter | Not applicable, hosted only | | Cloudflare | Not applicable, hosted only |
2. Is the thing you need on the free side of that line today, and will it still be in a year? Governance and high availability are the features teams need second, not first, which is exactly why they are the ones most often gated.
3. Do you actually need self-hosting at all? If prompts traversing a vendor is fine, Cloudflare AI Gateway is free on every plan and OpenRouter has a catalogue nothing self-hosted matches. A lot of teams self-host out of habit rather than requirement, and pay for it in operational attention.
If you do have the constraint, the shortlist is short and this post is mostly about reading licence files carefully. That is less exciting than a feature grid, and considerably more useful.
Everstack's gateway is documented here, with side-by-side comparisons for most of the projects above.