Ship AI changes
without the guesswork.
Canary deployments for LLM prompts and models. Split traffic, evaluate quality with an AI judge, and auto-rollback when things break — before users notice.
Live Traffic Flow
How It Works
Three-step loop running automatically, 24/7.
Split Traffic
Route a percentage of requests to a new prompt or model version. Users see no difference — Repath transparently proxies both.
Evaluate Quality
An LLM judge scores every response on your criteria — helpfulness, accuracy, completeness. Async, never slows down requests.
Decide & Act
Every 30 seconds, the controller checks scores. Quality good? Advance to more traffic. Quality dropped? Instant rollback.
Progressive Rollout in Action
Traffic advances through gates automatically when quality holds
AI models break silently
You can't detect "responses got 23% worse" by watching error rates. Zero HTTP errors. Zero exceptions. Just degraded output your users silently abandon.
Quality Score Over Time
When the new prompt degrades, Repath catches it instantly
Provider Updates Break You
OpenAI and Anthropic push model updates without notice. Your carefully-tuned prompts degrade overnight — zero errors in logs.
Weeks Before You Notice
Without quality scoring, prompt regressions hide for days. By the time users complain, the damage is done.
Feature Flags Can't Help
Feature flags control deployment, not quality. They can't answer: 'is the new version producing better outputs?'
Everything You Need
Purpose-built for AI deployment safety. Not feature flags bolted onto LLMs.
Canary Deployments
5% → 25% → 50% → 100% with quality gates at each step. Configurable weights, durations, thresholds.
LLM-as-Judge
GPT-4o-mini scores every response. Define criteria in plain English. Fully async — never slows requests.
Auto-Rollback
Quality < threshold? Traffic returns to baseline in <500ms. No human intervention needed.
Sub-2ms Overhead
Rust gateway with lock-free config. 50K+ req/s per instance. No GC, no runtime overhead.
Audit Trail
Every advance/rollback decision logged with exact scores. Full visibility into why actions were taken.
Self-Hosted
Docker Compose — one command. Data stays on your infra. No vendor lock-in.
Drop-in Integration
Change base_url in your OpenAI client. One line. No SDK, no wrappers, no code changes.
Real-time Dashboard
Live traffic split, quality graphs, decision timeline. See everything as it happens.
YAML Config
Declare rollouts as code. Version control your deployment strategy. GitOps-ready.
Why Repath?
The only tool that auto-rolls back based on semantic quality — not just error rates.
Semantic quality evaluation
Not just error rates — we evaluate whether responses are actually good using LLM-as-judge scoring on your criteria.
One-line integration
Change base_url in your OpenAI client. That's the entire integration. No SDK, no wrapper functions.
Quality drives rollout decisions
Rollback triggers on 'responses got less helpful' — not just latency spikes or error rate increases.
Works with any LLM provider
OpenAI today. Anthropic and Gemini coming Q3 2026. Any OpenAI-compatible endpoint works now.
vs. Traditional Tools
Built for Production
Rust gateway, Python evaluators, Next.js dashboard. Not a prototype.
Roadmap
Where we are and what's coming. Building in public.
Core Platform
Multi-Provider & Cloud
Advanced Evals
- Drift detection
- Human feedback loops
- A/B with significance
- Custom judge models
Enterprise
- SSO / SAML
- Team RBAC
- Kubernetes Helm chart
- Enterprise SLA & support
Running in 60 Seconds
Clone, configure, run. No account needed.
Clone & configure
Start all services
Point your app
Create a canary rollout
Stop shipping AI blind.
Know if your prompt change is better or worse — before your users do.