Your AI agent spent $500 on OpenAI overnight and nobody noticed. AXME tracks cost per agent in real time with budget limits.
One runaway loop. One missing stop condition. One agent calling GPT-4 in a retry loop for 8 hours. Monday morning: a $500 invoice and no way to trace which agent, which task, or which prompt caused it.
Alpha - Built with AXME (AXP Intent Protocol). cloud.axme.ai - contact@axme.ai
AI agents call LLMs. LLMs cost money per token. Nobody tracks it.
Friday 5:00 PM:
Deploy "research-agent" to production
Agent processes customer tickets, calls GPT-4 for each one
Expected: 200 tickets/day, ~$8/day
Friday 11:00 PM:
Bug in ticket dedup - agent reprocesses same tickets in a loop
Each iteration: 4 LLM calls x $0.03 = $0.12
Loop runs 50x/hour
Saturday 3:00 AM:
Agent has made 12,000 LLM calls
Cost so far: $360
Nobody is watching
Monday 9:00 AM:
OpenAI billing alert (you set it at $500)
Total damage: $487
No logs showing which agent, which intent, which loop
What breaks:
- No per-agent cost tracking - OpenAI shows total org spend, not per-agent
- No real-time visibility - you find out days later from the invoice
- No budget limits - nothing stops an agent from spending indefinitely
- No cost-per-intent - you know the total but not which task caused it
- No alerts - the $500 billing alert is too late, too coarse
This is not hypothetical. Every team running agents in production has a cost horror story.
AXME agents report cost as part of their regular heartbeat. The gateway aggregates it per agent, per intent, per time window. Cost policies enforce hard limits.
Agent heartbeat (every 30s):
{
"agent": "agent://myorg/production/research-agent",
"intent_id": "int_abc123",
"status": "processing",
"metrics": {
"cost_usd": 0.03, <-- cost since last heartbeat
"llm_calls": 4,
"tokens_in": 2100,
"tokens_out": 850
}
}
Gateway receives heartbeat:
1. Accumulate cost for this agent -> $12.47 today
2. Accumulate cost for this intent -> $0.87 for this task
3. Check against cost policy -> limit: $50/day
4. Under limit? Continue.
5. Over limit? Pause agent, notify owner.
No separate billing integration. No scraping OpenAI dashboards. Cost data flows through the same channel as agent status.
pip install axme
export AXME_API_KEY="your-key" # Get one: axme loginfrom axme import AxmeClient, AxmeClientConfig
import os
client = AxmeClient(AxmeClientConfig(api_key=os.environ["AXME_API_KEY"]))
# After each LLM call, report the cost
def call_llm(prompt: str) -> str:
response = openai.chat.completions.create(
model="gpt-4",
messages=[{"role": "user", "content": prompt}],
)
# Extract cost from usage
tokens_in = response.usage.prompt_tokens
tokens_out = response.usage.completion_tokens
cost_usd = (tokens_in * 0.03 + tokens_out * 0.06) / 1000
# Report to AXME mesh
client.mesh.report_metric(cost_usd=cost_usd)
return response.choices[0].message.content# Set budget limit via API
curl -X PUT https://api.cloud.axme.ai/v1/mesh/agents/{address_id}/policies/cost \
-H "x-api-key: $AXME_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"max_intents_per_day": 500,
"max_cost_per_day_usd": 50.00,
"max_intents_per_hour": 100,
"action_on_breach": "block"
}'Or set it from the dashboard at mesh.axme.ai - select agent, set cost policy, save.
When the agent exceeds $50/day, the gateway returns HTTP 429 on new intents. No notifications (yet) - check the dashboard.
View cost data at mesh.axme.ai with per-agent breakdown. Day, week, month views.
+-----------+ heartbeat +----------------+ dashboard +-----------+
| | (cost_usd) | | (aggregate) | |
| Agent | -------------> | AXME Cloud | ------------> | mesh |
| | | (gateway) | | .axme.ai |
| | pause/kill | | alerts | |
| | <------------- | cost policy | ------------> | email |
| | if over limit | enforcement | if threshold | slack |
+-----------+ +----------------+ +-----------+
- Agent makes LLM calls and tracks token usage locally
- Every heartbeat (30s), agent reports
cost_usdto AXME gateway - Gateway accumulates cost per agent per day/hour window
- Gateway checks cost policies on every new intent
- If limit exceeded, gateway returns HTTP 429 (
action_on_breach: "block") - Dashboard shows real-time cost per agent with day/week/month views
# Your agent code - no cost tracking
def process_ticket(ticket):
response = openai.chat.completions.create(
model="gpt-4",
messages=[{"role": "user", "content": ticket["body"]}],
)
return response.choices[0].message.content
# How much did this cost? Check OpenAI dashboard... in 24 hours.
# Which agent? Which task? No idea.
# Budget limit? Set a billing alert at $500 and hope for the best.# Same agent code + AXME cost reporting
def process_ticket(ticket):
response = openai.chat.completions.create(
model="gpt-4",
messages=[{"role": "user", "content": ticket["body"]}],
)
# 2 lines: calculate and report
cost = (response.usage.prompt_tokens * 0.03 + response.usage.completion_tokens * 0.06) / 1000
client.mesh.report_metric(cost_usd=cost)
return response.choices[0].message.content
# Cost visible in real time. Per agent. Per intent.
# Budget exceeded? Agent pauses automatically.- AXME - project overview
- AXP Spec - open Intent Protocol specification
- AXME Examples - 20+ runnable examples across 5 languages
- AXME CLI - manage intents, agents, scenarios from the terminal
- AI Agent Checkpoint and Resume - durable agent execution
- Async Human Approval for AI Agents - async approval with reminders
MIT - see LICENSE
Built with AXME (AXP Intent Protocol).