Kill your agents before they kill your budget.

tollgate — checkout-bot
$ tollgate init checkout-bot → issuing per-agent token · $30/mo ceiling → compiling budget policy… ✓ ready — request boundary armed

For every agent to hold its own token and its own monthly ceiling, never a shared key.

To check spend before each request forwards to OpenAI or Anthropic, not after the invoice.

Frozen the moment a loop runs away — one agent, or a batch, never the whole fleet.

Every agent gets its own token and its own ceiling.

Agents call Tollgate instead of calling OpenAI or Anthropic directly. Every request is checked against that agent's budget before a dollar leaves. Overspend is prevented, not reported back to you a month later.

Request inspectedLive
POST /openai/v1/chat/completions
agentcheckout-bot
assistantopenai
modelgpt-4o
tokens0
spend / budget$0.00 / $0.30
Checking… 
agent-scraper · shared key sk-…4f2
→ POST /v1/chat/completions 200 1,024 tok → POST /v1/chat/completions 200 1,024 tok → POST /v1/chat/completions 200 1,024 tok
   … retry loop ×3,000
spent  $3,912
budget none
alert  none

Stop one agent without stopping the fleet.

Pulling a shared API key takes down every agent using it. Tollgate freezes the one that is misbehaving, and the next request it sends comes back denied.

One agent, or a batch

Select any number of agents and freeze them together. The rest of the fleet keeps running on its own tokens.

Soft or hard

A soft freeze lets in-flight work finish. A hard freeze denies the next request outright. Both are one click.

No restart, no rotation

Freezing flips a Redis flag the proxy already reads. There is no deploy, no key rotation, and no downtime for anything else.

Freeze event
→ operator: freeze agent-scraper → redis flag set ok · 0ms → next request 403 agent_frozen no restart · no propagation delay

Know what every agent spent, and what it was denied.

Cost is metered from the provider's own usage numbers, not estimated. Every request lands in an audit log with the agent, assistant, model, token count, and what it cost — including the ones that were turned away.

WhenAgentAssistantModelStatusTokensCost
14:02:11checkout-botopenaigpt-4oAllowed1,204$0.011
14:02:09research-botanthropicclaude-sonnet-4Allowed8,410$0.084
14:02:04agent-scraperopenaigpt-4oOver budget
14:01:58checkout-botopenaigpt-4oAllowed912$0.009

Enforced at the request boundary, not after the bill.

Enforcement point
Pre-forward
Freeze propagation
Next request
Overspend bound
≤ 1 call
Budget scope
Per agent / mo
Warning / cutoff
80% / 100%
Dashboard refresh
Every 3s
Providers
OpenAI · Anthropic
Control plane
Redis fast-path

Drive the real dashboard.

Not a mockup or a video. This is the same Agent Dashboard component the product ships, running against a simulated traffic feed. Start the traffic and watch a budget fill.

Demo feed0 reqs / 0.65s

Agent Dashboard

Your group of agents, all managed in one place.

All agents
Polling every 3s
NameAssistantStatusSpend / budgetTokensEdit
checkout-botchatgptActive
$0.00 / $0.30
x
0%
0
data-scraperchatgptActive
$0.46 / $0.50
x
92%
52,310
nightly-backfillchatgptStopped
$0.00 / $1.00
x
0%
0
research-agentclaudeActive
$0.41 / $2.00
x
21%
18,420

Put a meter on your agents.

For larger fleets, custom limits, or self-hosting, send us a note below.