Anthropic capped programmatic Claude per subscription on June 15. RelayPlane spreads load across multiple credentials, fails over before 429s, and stops runaway agents at the budget cap. One proxy, multi-provider, local by default. One env var, no code changes.
Anthropic capped programmatic Claude per subscription on June 15. RelayPlane spreads load across multiple credentials, fails over before 429s, and stops runaway agents at the budget cap. One proxy, multi-provider, local by default. One env var, no code changes.
Open localhost:4100 afterrelayplane start. Live cost, per-model spend, per-agent breakdown, and cost guardrails. No telemetry, no login. Pro unlocks the kill switch, daily cap, and routing mode from the same screen.
Programmatic Claude is now metered at full API rates on top of your subscription. The effective cost multiplier across published math runs 12x for light scripting up to 175x for heavy Sonnet fleets. RelayPlane reroutes the same workload across the credential pool, quota-aware fail-over, and cheap-model offload, so the bill stays inside the plan.
Multipliers from publicly posted June 15 math (Theo Browne, Magna Capax, HN discussion). RelayPlane band assumes credential pool spans Pro / Max / API keys with quota-aware fail-over enabled.
ANTHROPIC_BASE_URL=http://localhost:4100Open Claude Code issue #26171: an agent burned 72,900 tokens over 21 minutes in a thinking loop with zero useful output. Pre June 15 that was annoying. After, every minute of that loop is metered at full API rates. RelayPlane catches velocity spikes, token explosions, and repetition loops in a sliding 100-request window and stops the bill at the budget cap.
Tenant kill-switch HTTP endpoint is on the next-2-week roadmap. Anomaly detection and budget caps are live in the local proxy today.
The proxy classifies every request before it leaves your machine. Routine work goes to fast, cheap models. Judgment work goes to the frontier. You stop paying $15/Mtok for tasks that need $0.10.
Observe is in production. Govern is shipping in pieces (budget caps and anomaly detection today, tenant kill-switch endpoint next). Verify is on the roadmap and runs out of band today.
Every LLM request flows through the proxy with full attribution. Cost, model, task type, tokens, latency, all live, all namespaced. Per-agent breakdown uses the system-prompt fingerprint, no annotation work required.
Daily, hourly, and per-request budget caps with block, downgrade, warn, or alert actions. Velocity spikes, repetition loops, and token explosions are detected in a sliding window. Multi-tenant kill-switch endpoint is next on the roadmap.
Before an agent marks a task done, RelayPlane will score the diff and acceptance criteria with a cheap judge model. Failing tasks retry. Today this lives as a separate evaluator in the clawd pipeline. Moving it into RelayPlane is on deck.
The numbers below are real and current, not aspirations. The proxy version is the live npm tag, the stars are pulled from GitHub at build time, the providers are the ones we wire to in v1.9.38 today.
Works with Claude Code, Cursor, OpenClaw, and any agent that supports ANTHROPIC_BASE_URL or OPENAI_BASE_URL. Point your agent at localhost:4100 and you are done.