agentshield/content/hn-submit-ready.txt at main · ghzhost/agentshield · GitHub

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
HN POST READY, copy-paste this into https://news.ycombinator.com/submit

TITLE: AgentShield – A firewall for AI agent spending (stdlib, 56/56 eval)
URL: https://agentshield.fly.dev

STATUS: Already submitted! Post URL: https://news.ycombinator.com/item?id=49250917

POST BODY (if text post needed for future reference):
---
At 3 AM, an AI agent I built made 21 API calls in 60 seconds. $2,800 gone before I woke up.

Existing solutions failed:
- API rate limits: kicked in too late (per-minute, not per-dollar)
- Budget alerts: emailed me while I was asleep
- Manual monitoring: doesn't scale past 3 agents

So I built AgentShield, 5 composable spending rules evaluated per-transaction in under 1ms.

What it does:
- Per-transaction limit (max $500/single call)
- Daily total cap ($2,000/day per agent)
- Velocity detection (flags >10 txns in 60 minutes)
- Merchant allowlist (only approved API providers)
- Category blocking (no crypto exchanges, no adult content)

Tech: Pure Python 3.11 stdlib. No pip install. SQLite + WAL mode. SSE for real-time alerts. Multi-tenant isolation out of the box.

The eval gym has 56 labeled scenarios across 7 categories. Currently 56/56 on first run.

Edge cases:
- Amount exactly at limit → APPROVED (not strictly greater)
- $0.01 over limit → BLOCKED
- Missing amount field → FLAGGED
- Empty rules list → APPROVED

Try the risk calculator (no signup): https://agentshield.fly.dev/tools/risk-calculator/
Live eval results: https://agentshield.fly.dev/eval

Feedback wanted:
1. Rule composition model, does first-match-wins cover real-world agent spending?
2. What rules would you add for your own AI agents?
3. Worth adding Slack/Teams alert integration?