Cut your AI costs by 30-50%.
Without changing your code.
“I was spending $200/month on Claude API calls for my coding agent. After installing RelayPlane, it dropped to $87. Same quality output—it just stopped using Opus for simple tasks like file reads and status checks.”
Matt Turley
Founder, RelayPlane · 20 years building software
Yes, this is the founder's own story. Early user testimonials coming soon.
How intelligent routing actually works
The Swarm explained in 3 sentences
Analyze metadata
Looks at task type, token count, and complexity signals — never your actual prompts.
Check patterns
What worked for similar requests across all users? “Code review under 3K tokens succeeds 94% on Haiku.”
Pick optimal model
Chooses the cheapest model that meets your quality threshold — if you need 95% success, it won't risk 87%.
Network effect
Every user (free or paid) contributes anonymized success/fail data. More users = smarter routing for everyone. We're early—currently 50+ beta users contributing routing data.
What if routing picks the wrong model?
This is the fear everyone has. Here's exactly what happens.
Scenario: Haiku returns garbage
Worst case: You paid for both calls. Still cheaper than always using Opus.
Built-in fallback chains
Circuit breakers: If a model fails 5x in a row, we skip it automatically for 30 seconds.
The math still works: Even with occasional fallback calls, users save 30-50% because most simple tasks do succeed on cheaper models. You're not paying Opus prices for file reads.
Start saving in 60 seconds
Three steps. Zero code changes.
Install the proxy
npm install -g @relayplane/proxyMIT licensed. Runs 100% locally.
Point your agent at it
export ANTHROPIC_BASE_URL=localhost:3001One environment variable. No code changes.
Watch your costs drop
relayplane statsReal-time dashboard shows exactly what you're saving.
More than just a proxy
RelayPlane is a complete agent operations platform. Here's what's built and ready.
Policy Engine
Complete control over which models can be used, when, and by whom. Set allowlists, blocklists, cost thresholds, and approval requirements. “Never use GPT-4 for customer data” — enforced automatically.
Included in Max tier
Compliance & Audit
Decision Audit Trail
Know exactly why each call was routed where it was. Full explainability for compliance.
Approval Gates
Require human approval for expensive operations. Multi-approver workflows.
Operations & Cost Control
Budget Controls
Hard limits by day/week/month. Get alerts before you blow through spend.
Pattern Detection
Spots anomalies, cost spikes, and failure patterns. Suggests optimizations.
Team Collaboration
Teams & Workspaces
Multi-user access with role-based permissions. Shared learnings across team agents. Separate budgets per team.
Full dashboard UI included (53 components). This isn't vaporware.
See documentation →Free forever. Pro when you're ready.
7-day free trial on Pro — no credit card required.
Free
For side projects and experimentation.
- Local proxy
- Static routing rules
- Local learning
- 30% average savings
- MIT licensed
- Offline mode
Break-even math: Pro pays for itself at $60/mo API spend. If you're spending $100+/mo, you'll save $20-50/mo after the subscription.
We never see your prompts.
Your data stays on your machine. We only see anonymized metadata for routing decisions.
What we collect
Anonymized metadata only
What we NEVER collect
Zero access to your content
Common questions
Technical
What if the cheap model returns garbage?
RelayPlane automatically retries with a better model. You pay for both calls, but that's still cheaper than always using Opus. The Swarm learns from failures so routing improves over time.
How much latency does this add?
Local proxy: <5ms. Swarm API call: 50-100ms. For a 2-5 second LLM call, this is negligible. If latency matters more than cost, configure more aggressive model selection.
What if the Swarm API goes down?
Proxy falls back to static rules automatically. Your agent keeps working. This is baked into the architecture—we assume network calls can fail.
Trust & Business
Can I audit exactly what's being sent?
Yes. Run `relayplane-proxy --audit` to see every payload before it's sent. Or read the MIT-licensed source code on GitHub.
What if you shut down?
The local proxy works forever—it's MIT licensed software on your machine. Only the Swarm API would stop. You'd keep ~30% savings on static rules.
Will this break my agent?
No. If RelayPlane can't route, it passes through to your default model. Worst case: you pay what you would have paid anyway.
How do I know I'm actually saving money?
Run `relayplane stats` or check the cloud dashboard. We show you exactly how much you've saved vs. what you would have spent.
Built by a real person.
Matt Turley
Solo Founder · 20 years in software
I built RelayPlane because I was spending $200/mo on Claude API calls and realized most of them didn't need Opus. Now I spend $87.
This is a real product from a real person, not a VC-funded growth hack. I've been independent for 16 years and I'll be around to fix bugs and answer questions.