OpenRouter alternatives
that are not another 400-model table
Eden AI (3 Aug) and TrueFoundry (7 Aug) both rank this query by putting themselves first. Live sites say otherwise: LiteLLM is the self-host proxy, Requesty is the Frankfurt router with PII scrubbing, Eden is the French multimodal aggregator. Copperway is the EU PII vault — not a 500-model table.
Direct answer
Two roundups own this SERP this week. Eden AI (3 Aug 2026) lists ten tools and ranks itself #1 as a French GDPR-native multimodal API. TrueFoundry (7 Aug) lists five and ranks itself #1 as a US VPC gateway. Checked live: if you want a pip-install proxy, that is LiteLLM. If you want a managed Frankfurt router with PII scrubbing, that is Requesty (UK company, AWS eu-central-1). If you want OCR/speech plus LLMs from Lyon, that is Eden AI — still forwarding to Bedrock, Vertex, and Azure in the EU. If you need a PII vault on an EU-incorporated path, try Copperway Playground. Residency is not sovereignty — Eden’s own article says so.
The LLM gateway keyword war (what we will not copy)
OpenRouter, LiteLLM, Portkey, Cloudflare AI Gateway, and a long tail of comparison blogs are fighting four buckets. AGICY enters only the buckets we can fulfil without lying.
- Category: LLM gateway, AI gateway, unified LLM API, OpenAI compatible proxy. Cloudflare and LiteLLM own raw “AI gateway” domain authority. We qualify: EU AI gateway, OpenAI compatible EU gateway.
- Pain: automatic fallback and retry, semantic caching, track API spend by user. Portkey, Orq.ai, and Helicone publish the code snippets. Copperway does semantic compression (SCE) and a PII vault — not a vector cache and not a retry mesh.
- Poaching: OpenRouter alternatives, LiteLLM vs Portkey, OpenRouter vs LiteLLM. This page is that matrix. One hub — not hundreds of AIML-style model doorways.
- BYOK long-tail: Cline, Cursor, Vercel AI SDK. Same move as LiteLLM: keep the OpenAI spec. Set
baseURLto Copperway. No fake one-click Cline button.
Eden AI’s top 10 — checked on the live sites
Eden’s table scores catalog size and “EU data residency.” We opened each homepage. Several rows are stale or self-serving.
| Eden rank | What the live site actually is | AGICY note |
|---|---|---|
| 1. Eden AI | Lyon, France. 500+ models including OCR, speech, vision, translation. ZDR, SOC 2, ISO 27001 on the homepage. EU endpoint api.eu.edenai.run. Fallback. Their own EU table lists Bedrock Ireland, Vertex EU, Azure EU Data Zone, Mistral, OVH, Scaleway. | Closest SERP rival for “European OpenRouter alternative.” French HQ is real. Inference still sits on US-parented clouds unless you pick Mistral/OVH/Scaleway. Do not copy the 500-model multimodal catalog. |
| 2. Portkey | Homepage now says Portkey is PRISMA AIRS AI Gateway, generally available, Palo Alto Networks. SF address. 1,600+ LLMs, PII redaction, caching, MCP gateway, RBAC. | Eden still describes an independent startup. It is a US security vendor. CLOUD Act applies to the operator. |
| 3. Requesty | Not “lightweight.” Live: 600+ models, 5% markup, failover, caching, Cursor/Cline OAuth, router.eu.requesty.ai, PII scrubbing, ZDR. Security page: Frankfurt, GDPR. SOC 2 Type II is claimed in the FAQ and “in progress Q3 2026” on the same page — treat as unverified. | Closest product clone: OpenAI baseURL + EU region + PII mask. UK company on AWS Frankfurt. Same CLOUD Act test as every other US/UK router with an eu-central-1 flag. |
| 4. Helicone | Observability-first gateway. Live homepage: joining Mintlify. OSS Docker/Helm still exists. | Spend dashboards, not jurisdiction. Do not clone. |
| 5. LiteLLM | BerriAI, SF. 140+ providers, ~1,900 models, 53k GitHub stars, 240M+ Docker pulls. Self-host / air-gap. Semantic cache, prompt compression, PII guardrails, virtual keys. Rust gateway claims 0.66 ms p99 overhead. | Eden’s “100+ models, limited observability” is undersold. This is the self-host winner. Copperway is managed, not a Helm chart. |
| 6. Cloudflare AI Gateway | Docs (updated Apr 2026): analytics, cache, rate limit, retry, model fallback. Free on all plans. Not a model catalog — you bring providers. | Domain-authority trap. Do not fight “AI Gateway” as a raw keyword. |
| 7. Together AI | Own GPU cloud: open-weight inference, fine-tune, H100/B200 clusters. OpenAI-compatible. Not a multi-provider router. SF. | Different category (neocloud). Point cost buyers at cheapest EU inference and planned Cyprus leasing. |
| 8. NanoGPT | Hobby marketplace / chat UI. Wide catalog, crypto, TEE claims. Eden’s own vs-table contradicts 979 vs 400+ models. Not a production gateway. | Catalog gravity for individuals. Not our buyer. |
| 9. Vercel AI Gateway | Hundreds of models, 0% markup, failover, text/image/video/audio/embeddings/rerank, vercel ai-gateway coding-agents setup for Cline/Claude Code/Codex. ZDR is a routing policy to ZDR providers. US company. Team-wide ZDR is a paid add-on in Eden’s table. | This is OpenRouter’s “every modality + BYOK coding agents” play with no fee. We already refused to clone that catalog. Keep the baseURL snippet only. |
| 10. Kong AI Gateway | Govern LLM, MCP, and A2A on Kong. PII sanitization, semantic cache, semantic routing, token quotas. You configure providers. US control plane; data planes can sit in your EU VPC. | Enterprise API platform, not a try-now playground. |
Eden’s “Best European OpenRouter Alternatives” section is the honest part of their article: residency ≠ sovereignty; Frankfurt does not make a UK or US company European; an EU gateway that forwards to a US-hosted model is not EU inference. Apply that test to Eden too — Bedrock Ireland is still AWS. Copperway’s answer is PII stripped before upstream plus an EU-incorporated operator, then planned Cyprus compute via SRA. Detail: EU alternative to OpenRouter.
Comparison matrix (failure mode, not score /25)
| If you need | Winner on that SERP | AGICY move |
|---|---|---|
| French HQ + multimodal + EU endpoint | Eden AI | Closest EU SERP rival. Still Bedrock/Vertex/Azure behind the EU flag. Do not copy 500 expert models. |
| Managed Frankfurt router + PII scrub + 5% fee | Requesty | Closest product clone. UK company, AWS eu-central-1. Copperway converts on jurisdiction, not markup. |
| Gateway in your VPC, RBAC, tracing | TrueFoundry / LiteLLM / Kong | Self-host or US enterprise control plane. Not a live Copperway installable. |
| 0% markup, every modality, coding-agent CLI | Vercel AI Gateway | Catalog + DX fight. Do not clone. |
| Self-host, zero middleman fee, docker-compose | LiteLLM | You operate the proxy. Copperway is managed EU routing, not a README you run. |
| Fallback, retry, semantic cache, guardrails | Portkey / Orq.ai | Honest no. SCE compression cuts tokens; it is not semantic caching. |
| Track API spend by user / multi-tenant keys | Helicone / Portkey | Not a Helicone clone. SRA is reserved capacity. |
| Evals and experiment tracing | Braintrust | Different product. Do not rank as a gateway. |
| PII vault + EU path + OpenAI SDK drop-in | Copperway | Try free AI · Gateway · Reserve SRA |
OpenRouter’s known soft spots on these SERPs are the ~5.5% credit card fee, thin compliance certificates, and weak enterprise monitoring. LiteLLM attacks the fee. Portkey attacks the ops gap. AGICY attacks jurisdiction: a US marketplace with an eu.openrouter.ai flag is still a US operator.
LiteLLM vs OpenRouter — TrueFoundry’s 12 Aug fork
TrueFoundry’s LiteLLM vs OpenRouter (published 12 Aug 2026) is the poaching page for that exact query. The useful fork is real. The third-place finish they award themselves is sales copy. We do not repeat their 3–4 ms / 10 ms / 350 RPS numbers, and we do not repeat “LiteLLM is high-latency and only for prototypes” — that is TrueFoundry attacking a 53k-star proxy, not a measurement we ran.
| LiteLLM | OpenRouter | TrueFoundry (self-rank) | Copperway | |
|---|---|---|---|---|
| What it is | Open-source proxy + Python SDK you run (Docker / Helm / air-gap) | Managed SaaS aggregator. Unified credits. No self-host option (their article) | Managed gateway pods in your VPC. Ensemble Labs Inc, San Francisco | Managed EU OpenAI-compatible path with a PII vault |
| Catalog | TrueFoundry writes “100+ models.” Live LiteLLM: 140+ providers, ~1,900 models. You bring keys. | Hundreds of models, one invoice, ~5.5% card fee. Auto failover and spend-index routing | 250+ public models plus vLLM / TGI you host | Not a 400-model table. Playground + planned Cyprus capacity |
| Ops you own | Deploy, scale, patch, Redis, Helm. Fallback and budgets in YAML | None. You consume an edge endpoint | They claim you keep control and lose the babysitting. Still a US control plane | None for the proxy. You do not get LiteLLM’s GitOps YAML |
| Can you compose them? | Yes: LiteLLM can take OpenRouter as a provider. Governance in your VPC, catalog in OpenRouter’s cloud. Prompts still leave the EU if that is where OpenRouter routes. | Positions itself as the layer above both | Not a LiteLLM wrapper. Set baseURL once | |
| CLOUD Act / EU | Wherever you host the proxy. BerriAI is SF. Self-host in Frankfurt is residency of the box, not of every upstream | US operator. An eu.openrouter.ai flag is still OpenRouter | US company. VPC in eu-central-1 is not CLOUD Act-proof | EU-incorporated operator. PII stripped before upstream. Campus is pre-construction |
If the search was “LiteLLM vs OpenRouter,” pick on failure mode: self-host and skip the middleman → LiteLLM. One credit pool and no cluster to run → OpenRouter. Gateway in a VPC you already operate → TrueFoundry or Kong. OpenAI SDK + PII vault on an EU path → Copperway Playground. Do not install Copperway as a Helm chart, and do not expect OpenRouter’s catalog.
Unified LLM API — the only category fight we join
Developers will not rewrite the app. LiteLLM and OpenRouter both win “call Anthropic using OpenAI format.” Copperway uses the same hook: one OpenAI-compatible endpoint, PII masked before upstream.
from openai import OpenAI
client = OpenAI(
base_url="https://proxy.agicy.com/v1",
api_key="YOUR_API_SECRET",
)
response = client.chat.completions.create(
model="gpt-4o-mini",
messages=[{"role": "user", "content": "Summarise this clause."}],
stream=False,
)Same pattern in Node and in the Vercel AI SDK: swap baseURL, keep the client. Full snippets live on Copperway → Integration. That is the “OpenAI compatible proxy” query without pretending we are LiteLLM’s GitHub docs.
Pain keywords we refuse to fake
When an OpenAI region dies, people search LLM automatic fallback and retry. When the bill explodes, they search semantic caching for LLMs and track API spend by user OpenAI. Those pages convert because the product actually retries, caches, or meters. Copperway today:
- Does: PII vault, EU routing, SCE compression, OpenAI-compatible chat completions, free playground with daily limits.
- Does not: Portkey-class failover meshes, vector semantic cache, Helicone-style per-user spend dashboards, Kong semantic routing, OpenRouter spend-index auto router.
Semantic routing (“send math to model A”) is Kong’s fight. Our routing story is the PII vault + EU path — see secure LLM gateway.
Which page to open next
- Developers who typed OpenRouter alternatives, LiteLLM vs OpenRouter, or unified LLM API → Copperway Playground.
- DPOs / CTOs who typed CLOUD Act or European alternative → EU alternative to OpenRouter.
- Offtake → SRA and GPU leasing Cyprus. The campus is pre-construction.
We will not generate a landing page per open-weight SKU. Catalog gravity lives on the model catalog / leaderboard. This page owns the comparison query.