Set a ceiling on agent spend.

Ceiling one — the balance itself

Credits are prepaid. When they are gone the requests stop, and there is no invoice arriving afterwards for the part you did not notice.

Ceiling two — per-key caps

Four limits are enforced at the edge, per key. Open reservations count against them the moment a request starts, so a burst of concurrent calls cannot slip past a cap that a serial one would have hit.

Defaults are lower for accountless keys and every one of them can be raised per key. When a request is refused the 429 carries Retry-After and names which limit was reached and where to raise it — see /docs/rate-limits.

What to actually set the cap to

Cost one turn, multiply by the turns you expect in a day, then set the cap somewhere above that and below what you would mind losing. The example below uses current catalog rates; replace its token counts and turn frequency with your own workload.

Example input per turn

4K tokens

System prompt, tool schema and recent history, re-sent each turn.

Example output per turn

250 tokens

A short reply or a tool call, not an essay.

Example runaway

480 turns

One turn a minute for 8 hours — an overnight loop nobody was awake for.

One turn costs $0.0004 on DeepSeek V4 Flash 0731 and $0.0012 on GPT 5.6 Luna. The useful reading of that spread is not which model is cheaper — it is that the same cap protects you very differently depending on the model behind it, so set the cap after you have chosen the model, not before.

Round up, not down. A cap that trips during normal operation trains you to raise it without thinking, which is how caps stop working. Set it at several times your expected day so that hitting it is genuinely information.

Ceiling three — one key per agent

The only one of the three that is a habit rather than a mechanism, and the one that does the most work.

Give each agent its own key and limit. If one hits its cap, the others keep working and you know which agent to inspect.

What hitting a ceiling looks like

Your agent's errors get read inside a chat app where we have no UI, so each one has to be self-contained.

curl https://api.minirouter.sh/v1/chat/completions \
  -H "Authorization: Bearer $MINIROUTER_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model":"meta/llama-3.3-70b","messages":[{"role":"user","content":"ping"}]}' \
  -D - -o /dev/null

# every response carries the balance, so an agent can warn you
# before it runs dry rather than after:
#   x-minirouter-balance-usd: 4.81
#   x-minirouter-cost-usd: 0.0004
#   x-minirouter-provider: together
#   x-minirouter-ttft-ms: 240