A local LLM proxy that shows you what every agent run costs and caps the runaway before it drains your budget. Routing each request to the model that fits is the mechanism that keeps the bill down.
RelayPlane routes the agent pipeline that ships this product. Every run behind this site is metered at list price, capped by the same budget guards, and written to a local ledger. We watch our own bill through it every day.
Open Claude Code issue #26171: an agent burned 72,900 tokens over 21 minutes in a thinking loop with zero useful output. You do not notice until the bill. RelayPlane watches a sliding window of your traffic and catches velocity spikes, token explosions, and repetition loops before they run up a charge, then stops at the budget cap.
Anomaly detection and budget caps are live in the local proxy today. All of it runs on your machine, with no account.
Put relayplane run --label nightly-backfill -- ./orchestrate.sh in front of any command. No SDK, no instrumentation, nothing inside your agent. The proxy already sees every request, so tagging the process at dispatch time turns the rollup into a group-by: cost per run, per agent, per thread, per retry.
$ relayplane run --label nightly-backfill -- ./orchestrate.sh Run nightly-backfill-20260905-a1b2c3 (nightly-backfill) completed 18m 42s cost $6.41 notional (all-opus baseline $19.02) requests 214 retries $1.10 (17%) band in [4.00, 9.00] 429s 0 model drift 3 agents: coder $3.88 (121 req, sonnet x118, haiku x3) | architect $1.94 (34 req, opus x34) dashboard http://localhost:4100/dashboard#run=nightly-backfill-20260905-a1b2c3
Every figure is notional list price for the traffic, never an invoice. The ledger is SQLite on your machine and nothing about a run is uploaded.
Open localhost:4100 afterrelayplane start. Live spend for the day, per-model and per-agent breakdown, and the full request ledger. Or stay in the terminal: relayplane watch is a live cost ticker. Nothing leaves your machine and there is no login.
Tell RelayPlane which tier of model handles simple, moderate, and complex work, or pin a specific model per agent. Every tool that speaks the OpenAI or Anthropic API goes through the same endpoint, so changing models never means changing code.
We run our own autonomous engineering pipeline through RelayPlane. Last week we flipped its coder role from Claude to Kimi K3 on OpenRouter by editing two fields in one JSON file, watched the live cost pane, and flipped back nine hours later when the numbers said no. That is the whole workflow.
// one role, one edit, no code changes "coder": { "provider": "anthropic", "model": "claude-sonnet-5" } // vs "coder": { "provider": "opencode", "model": "openrouter/moonshotai/kimi-k3" }
Observe is in production. Govern is shipping in pieces (budget caps and anomaly detection today, a 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. The per-agent breakdown uses the system-prompt fingerprint, so there is 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. A 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 our pipeline. Moving it into RelayPlane is on deck.
These are real numbers, not aspirations. The version is the live npm tag. The rest come from an autonomous engineering pipeline we run through this proxy every day: it routes, prices, and guards all of its traffic.
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.
