OpenRouter guardrails – Tokenminning

OpenRouter guardrails cap spend and restrict model access at the org and workspace level — before runaway agent loops hit your credit balance.

Hub: Tokenminning with OpenRouter.

This page answers openrouter spending limit, openrouter budget, and openrouter guardrails.

Workspace budgets

Workspaces isolate projects, teams, and agents. Each workspace can have a spending budget with automatic enforcement when the cap is reached.

Workspace	Purpose	Budget pattern
`dev`	Cline / Aider local	Low monthly cap
`ci`	Eval scripts, bots	Hard cap, haiku-only
`prod`	App traffic	Separate from coding

Pass workspace_id when querying the Analytics API to attribute spend per workspace.

Org guardrails

Guardrails control spending and model access for your organization:

Guardrail type	What it blocks or limits
Model allowlist	Frontier slugs on CI keys
Spending limits	Per-key and org-level caps
Sensitive info	API keys in prompts before they reach providers
Prompt injection	Regex patterns on inbound requests

Configure guardrails in the OpenRouter dashboard under your org settings. Pair with per-key limits — guardrails are the org-wide backstop.

Per-key spending limits

At key creation, set a spending limit. When the key hits the cap, requests fail fast instead of draining the shared credit pool.


Dev key   → $20/mo cap, haiku + sonnet allowlist
CI key    → $5/mo cap, haiku-only
Prod key  → separate pool, no coding-agent traffic

Rotate keys when a tool leaks into the wrong environment. Never commit keys — use env vars and secret scanners.

Notifications

Notifications alert before credits deplete or when a workspace approaches its budget. Route alerts to Slack or email so spikes surface during business hours, not Monday morning.

BYOK tradeoff

Bring your own key (BYOK) routes through your provider credentials. You pay provider rates directly; OpenRouter charges a smaller platform fee. Useful when you already have enterprise Anthropic or OpenAI contracts — but you lose unified credit pooling across providers.

Batch workloads

The Batch API runs inference asynchronously. Submit large eval or offline workloads as batches instead of synchronous agent loops — lower effective cost and no interactive latency pressure.


curl https://openrouter.ai/api/v1/batches \
  -H "Authorization: Bearer $OPENROUTER_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"input_file_id": "file-abc123", "endpoint": "/v1/chat/completions"}'

Reserve synchronous /chat/completions for interactive coding agents.

Agent SDK loops

The Agent SDK (callModel) runs multi-turn tool loops in one invocation. Each tool call re-bills. Cap maxSteps in your agent config and default to haiku-tier models for tool-selection turns.

Spending guardrails checklist

Separate API keys per tool and environment
Spending limit on every key
Workspace budgets for prod vs dev
Model allowlist on CI keys
Notifications before credit depletion
Batch API for offline eval, not agent loops

Troubleshooting

Key blocked mid-sprint — cap too low for legitimate traffic. Raise cap on dev key; keep CI key tight.

Guardrail rejected a valid request — check model allowlist and sensitive-info patterns. Allowlist safe internal domains if web-fetch tools are enabled.

BYOK bill higher than expected — BYOK removes OpenRouter credit pooling but not upstream provider rates. Activity still shows per-request cost.