RelayPlane sits between your orchestrator and your agents. Every request verified against spec. Every tenant isolated. Every dollar tracked. One API call to halt any tenant instantly.
| Scenario | Model mix | Weekly cost |
|---|---|---|
| Without routing | 100% Opus | $33.75 |
| With RelayPlane routing | 30% Opus · 70% Sonnet | $11.23 |
| You save per week | $22.52 (67%) | |
These are illustrative numbers based on a 30% Opus / 70% Sonnet routing estimate. Opus: $15/M tokens · Sonnet: $3/M tokens (Anthropic April 2026 pricing). Your actual savings depend on usage patterns.
When you run an agent swarm commercially — for clients, for production systems, for anything that matters — you need more than cost tracking. You need provable isolation, verifiable output, and a stop button that works in under one second.
Without a governance layer: Client A's runaway agent bleeds into Client B's budget. Agents mark tasks done without meeting spec. A billing emergency has no kill-switch. Auditors ask for logs you don't have.
*Default routes simple work to Sonnet. With provider keys configured, RelayPlane can route simple tasks to any cheaper capable model you choose.
Enter your monthly API spend
Matt Turley runs Continuum, a commercial agent swarm service. Every client gets their own tenant lane through RelayPlane. Requests from Client A are physically isolated from Client B at the proxy layer — not just logically separated in application code.
Every coder agent task includes its acceptance criteria. When the agent finishes, RelayPlane's spec-match plugin verifies the output before it reaches the verifier. Only passing tasks proceed. Failing tasks retry with escalation — automatically.
When a billing anomaly spikes unexpectedly, Matt calls DELETE /v1/tenants/acme-corp/kill. That tenant's traffic stops within the next request cycle — under one second. No waiting for rate limits or budget ceilings.
When Anthropic stopped supporting Max subscriptions in third-party tools, developers moved to the API and started paying attention to costs. But cost control was just the first layer. Now teams are running production agent swarms for clients and need the full governance stack: isolation, verification, audit, kill-switch.
RelayPlane started as a cost-routing proxy. It is now the governance gateway every commercial agent deployment needs.
Every LLM request flows through RelayPlane with full tenant attribution. Cost per request, model used, task type, tokens consumed, latency: all tracked and namespaced by tenant. Per-agent cost tracking identifies each agent by system prompt fingerprint. Tenant dashboards show spend, request volume, and top models — isolated from other tenants. Every trace linked to the tenant that generated it.
A layered policy engine with per-tenant enforcement. Hard budget caps block requests when daily limits are hit — no soft warnings, no grace periods unless you configure them. Tenant isolation ensures a runaway agent on one tenant cannot affect others. The kill-switch halts all traffic for a tenant within one request cycle. Anomaly detection catches runaway loops and cost spikes in real time.
Before any agent marks a task complete, RelayPlane's spec-match plugin evaluates the output against the task's acceptance criteria. The orchestrator POSTs the diff + criteria to /v1/spec-match. A cheap judge model (Haiku by default) evaluates each criterion and returns a structured pass/fail result with per-criterion evidence and confidence scores. Only pass: true results proceed. Failing tasks automatically retry.
Every action in RelayPlane is recorded in a tamper-proof audit chain. Each entry is checksummed and linked to the previous one — if anything is modified, verification fails. Export a compliance bundle for any tenant, any time range, in JSON, CSV, or JSONL.
Enterprise tiers retain audit logs for up to 7 years. Free tier: 7 days. Built on the same audit infrastructure used for SOC 2 evidence collection.
npm install -g @relayplane/proxy && relayplane init && relayplane startWorks with Claude Code, Cursor, OpenClaw, and any agent that supports ANTHROPIC_BASE_URL or OPENAI_BASE_URL.
Point your agent at localhost:4100 via ANTHROPIC_BASE_URL or OPENAI_BASE_URL. No risk: if RelayPlane goes down, your agent keeps working.
Run relayplane stats in your terminal for a quick cost summary.
Supports: Anthropic · OpenAI · Google Gemini · xAI/Grok · OpenRouter · DeepSeek · Groq · Mistral · Together · Fireworks · Perplexity
We learned this the hard way. Early versions hijacked provider URLs. One crash took everything down for 8 hours. Never again.
RelayPlane uses a circuit breaker architecture. After 3 failures, all traffic bypasses the proxy automatically. Your agent talks directly to the provider. When RelayPlane recovers, traffic resumes. No manual intervention.
"I was mass spending $200+/month running an agent swarm and had zero visibility into where the money was going. Turns out 73% of my requests were using Opus for tasks Haiku could handle."Matt Turley, Continuum
No. RelayPlane uses a circuit breaker architecture. If the proxy fails for any reason, all traffic automatically bypasses it and goes directly to your LLM provider. Your agent doesn't even notice. If RelayPlane can't route, it passes through to your default model. Worst case: you pay what you would have paid anyway. We learned this lesson the hard way and built the safety model first.
Yes. RelayPlane is open source (MIT license) and free to self-host. All features work locally with no account required. There are no paid tiers currently.
Telemetry is on by default. You can disable it (relayplane telemetry off). We collect anonymized metadata: task type label, token count, model used, latency, estimated cost, and an anonymous device ID. Your prompts, code, and responses are never collected. They go directly to LLM providers.
RelayPlane uses heuristic classification (token counts, keyword patterns, code block detection) to label requests by complexity: simple, moderate, or complex. When you enable routing, it maps these labels to models you choose, for example simple to Sonnet or a cheaper capable model, moderate to Sonnet, complex to Opus. Default behavior is passthrough unless you configure routing rules.
RelayPlane works with Anthropic API keys, pay-as-you-go or prepaid. Anthropic no longer supports Claude Max subscription tokens in third-party tools. If you're moving from Max to API keys, RelayPlane helps you control costs at the API level.
Yes. RelayPlane supports any OpenAI-compatible API: Claude, GPT-4o, Gemini, Mistral, open-source models via OpenRouter. The routing engine is model-agnostic.
It depends on your usage pattern, but most users see 40-70% cost reduction. The biggest savings come from routing simple tasks (which are typically 60-70% of all requests) to cheaper models. You'll see exact numbers in your dashboard within the first hour.
Different layer entirely. OpenRouter is a multi-provider gateway: you pick the model, it routes to the cheapest provider for that model. RelayPlane picks the right model for the task. RelayPlane is local-first (your machine, your data). OpenRouter is a cloud service you send all prompts through. They're complementary: you can use OpenRouter as a provider behind RelayPlane. RelayPlane adds cost tracking, task classification, and a local dashboard on top.
LiteLLM is a unified API adapter: call any provider with one SDK. RelayPlane does that and adds configurable task-aware routing, cost tracking, and a dashboard. LiteLLM requires code changes (import litellm). RelayPlane is a proxy: you set ANTHROPIC_BASE_URL to point at it. LiteLLM is a library you integrate. RelayPlane is infrastructure you deploy.
Two-layer pre-flight classification with zero latency overhead (no LLM calls). First, task type: regex pattern matching on the prompt across 9 categories (code generation, summarization, analysis, etc.), under 5ms. Second, complexity: structural signals like code blocks, token count, and multi-step instructions scored as simple, moderate, or complex. Routing rules map task type + complexity to a model tier. Cascade mode starts with the cheapest model and auto-escalates on uncertainty or refusal patterns.
Your prompts and responses go directly to LLM providers, never through RelayPlane servers. Telemetry (anonymous metadata: task type, token counts, model, cost) is on by default. Disable it with relayplane telemetry off. MIT licensed, fully auditable.
Nothing. Remove the proxy, your agents talk directly to providers again. No lock-in, no migration, no data hostage. It's MIT licensed. You can fork it and run your own if you want.
RelayPlane automatically retries with a better model. You pay for both calls, but that's still cheaper than always using Opus. Collective failure learning is on the roadmap. For now, the proxy logs these so you can adjust your routing rules.
The local proxy works forever. It's MIT licensed software on your machine. Only the mesh network would stop. You'd keep ~30% savings on static rules.
Run relayplane stats or check the dashboard. We show you exactly how much you've saved vs. what you would have spent.
Install RelayPlane. Tenant isolation, spec-match, kill-switch — all in 3 minutes.
npm install -g @relayplane/proxy && relayplane init && relayplane start