Compare

Honest, side-by-side.

All five projects share roots in the same problem space: route LLM traffic between providers. They diverge sharply on what they prioritize. Pick by fit, not loyalty.

Concern	kRouter	9router	OmniRoute	LiteLLM	OpenRouter
Cost & license
License	MIT	MIT	MIT	MIT	Closed source
Self-hosted (no data leaves your machine)					Cloud only
Cost of the tool itself	Free	Free	Free	Free	Per-request markup
Antigravity ban handling
Bootstrap metadata enum type	Numeric (matches binary)	String (triggers verify 403)	Numeric	n/a	n/a
Permanent-ban classifier	24h modelLock___all	Generic 1h cooldown
Test-Connection clears every lock
Quota tracking accuracy
Live per-model quota %		`\|\|0` paints exhausted as 100% used
Exhausted vs 0% remaining distinction	Amber 'Exhausted • resets in X'	Fake red 100% bar
TPM vs daily-quota disambiguation on 429	90s TPM cooldown	30m flat
Auto-extract wrapped Antigravity usage tokens	Auto-backfills historical 0/0 rows	All rows record 0/0
Routing & retry
Failover engine intelligence	Zenith AI scoring (latency + quota mathematical ranking)	Dumb fill-first sequential cascade	Sequential cascade	Priority sequential cascade	n/a
Per-provider concurrency limits	Kiro 4 · Claude 5 · Antigravity 2	Flat	Tunable	Tunable	Provider-side
Failover latency (RAM layer vs SQLite)	< 1ms (Sub-5ms HealthCache RAM layer)	~50ms disk lock per failure	~10ms edge state	~5ms Redis lock	n/a
Atomic backoff under concurrency	SQLite transaction	Race loses increments
MITM intercept (IDE traffic)
MITM stream-layer reliability	Bounds-checked frames, NGHTTP2 → HTTP/1.1 fallback	Stream errors as 'Truncated event message'	Different design (edge/Deno)
Antigravity / Kiro / Cursor IDE intercept
x-request-source anti-loop marker preserved
Provider features
Thinking config preservation through blacklist	Translates Claude/OpenAI shape	Blacklist strips, never runs
ZWJ obfuscator skips binary fields (images)		Corrupts inline_data → 400
Number of supported providers	89+	40+	35+	100+	100+
Security
SSRF guard on baseUrl (blocks cloud metadata)			Edge sandbox
Timing-safe CLI token compare
Per-IP brute-force lockout
Deployment
NPM package	@sifxprime/krouter	9router	—	litellm	—
Docker image	sifxprime/krouter	decolua/9router		ghcr.io/berriai/litellm
Update cadence	Hardening fixes 24-72h after report	Active feature dev	Active independent dev	Active	Active

Which is right for you?

The honest take.

kRouter

You want the hardening — verify-403 fix, MITM stream recovery, atomic backoff, SSRF guards — without paying for it.

9router

You want the upstream and don't mind hitting the bugs the fork has already patched. Fastest feature cadence.

OmniRoute

You prefer an edge/Deno deployment story and a different MITM stack. Comparable hardening philosophy.

LiteLLM

You need 100+ provider breadth and don't care about IDE intercept. Enterprise SDK pattern.

OpenRouter

You want zero ops, are fine paying a per-request markup, and don't need self-hosting.