Path is for teams that run Java and want their LLM traffic to stay inside their own JVM, database and process — not behind a Python-only sidecar or a vendor's SaaS.
- One process. A Spring Boot jar, Flyway migrations and an embedded H2 (or your Postgres). No Python runtime, no Node gateway, no sidecar to operate.
- Deterministic routing you can defend. Price and priority decide where a request goes; a customer dispute can be answered with "OpenAI returned 500, Groq was next".
- Billing that accounts for caches. Prompt, completion, cached and cache-write tokens are priced separately, so a cache-heavy workload is not overbilled 10×.
- Self-hosted by construction. Provider keys are AES-GCM encrypted at rest, admin tokens are scoped and audited, and nothing leaves your network except the upstream call.
- Boring operations. Readiness gates on seeding, shutdown drains long streams, per-provider circuit breakers, Prometheus, OpenTelemetry and an OpenAPI spec.
Be honest about this before choosing it.
| Path | LiteLLM | Portkey | OpenRouter | |
|---|---|---|---|---|
| Runtime | Java 21 / Spring Boot | Python | Node/Go, hosted | hosted SaaS |
| Model coverage | 2 wire formats, 4 seeded providers | 100+ providers | 250+ models | 300+ models |
| Hosted option | none (self-host only) | community | yes | yes |
| SDKs | any OpenAI SDK | Python SDK | JS/Python SDKs | JS/Python SDKs |
| Guardrails / PII | not built | plugins | built-in | built-in |
| Semantic cache | no (exact match only) | plugins | yes | no |
| Dashboard | JSON APIs + Postman/Bruno | UI | full UI | full UI |
| Licence | GPL-3.0 | MIT | Apache-2.0 core | proprietary |
Model count is the biggest gap. If you need Gemini's native protocol, Bedrock signing, or a hundred long-tail providers on day one, a larger project covers more ground.
Path's answer is not "more providers". It is "providers you already use, routed and billed precisely, in the runtime your team already operates".
No API keys needed — the demo talks to a mock upstream, so it shows the mechanics (routing, streaming, cache, metering) without spending anything.
bash tools/demo.shWhat it prints:
- the seeded catalogue (
/v1/models), including capability aliases - a non-streaming chat through
auto-fast, with provider and cost headers - a streaming chat, showing SSE chunks and
[DONE] - the same deterministic request twice, showing
X-Path-Cache: hitand zero cost /v1/usagefor the key: tokens, cache savings, recent calls
To point it at real providers instead, see the Quick start in the README
and set OPENAI_API_KEY, ANTHROPIC_API_KEY, GROQ_API_KEY or GEMINI_API_KEY.