Just open-sourced PurseString — open-source guardrails for AI agents that spend money.
It’s a self-hosted MCP server. Plug it into Claude, Cursor, OpenClaw, Hermes, etc.
How it works:• Agent wants to buy something and must give a reason• PurseString checks your rules (budget, merchant, amount, recurring)• Safe purchases go through on a single-use virtual card• Risky ones wait for human approval (agent can’t approve itself)• Full audit log + no double charges
Card providers: Lithic, https://t.co/LTtoTDK4sS & Airwallex (beta) + mock cards for testing.
“Let your agent buy, safely.”
https://t.co/kttN7gdlEX
Star it, try the mock rail, or open a PR.
Day 1/12: I entered the PayPal AI hackathon with PurseString, open-source guardrails for AI agents that spend money.
Building the PayPal rail in public: every purchase gets a policy check and a human OK before money moves.
Try it now with mock cards:
https://t.co/JA4nuEd3HT
#paypal #AIhackathon #OpenSource #AI #AIAgents
Small correction the card carries the intent ID, and the reason lives in the audit log under it. Refunds are logged against the original intent without editing it, each refund counts once, they can never add up to more than the charge, and the budget comes back. Unmatched refunds get flagged for a human
Open source: PurseString, guardrails for AI agents that spend money. 💳
Your agent shouldn’t hold your credit card. Instead:
• Must give a reason to buy
• Budget + merchant rules
• Single-use virtual cards
• Human approval for risky buys
• Audit log, no double charges
Self-hosted MCP server 👇
https://t.co/kttN7gdlEX
#opensource #AIAgents #MCP
@Shadowfetch ‘give a reason’ thats is exactly my favorite feature too, since it makes the audit log actually readable. thanks @Shadowfetch . really Would love your feedback!
#opensource#AIAgents#MCP
Just open-sourced PurseString — open-source guardrails for AI agents that spend money.
It’s a self-hosted MCP server. Plug it into Claude, Cursor, OpenClaw, Hermes, etc.
How it works:• Agent wants to buy something and must give a reason• PurseString checks your rules (budget, merchant, amount, recurring)• Safe purchases go through on a single-use virtual card• Risky ones wait for human approval (agent can’t approve itself)• Full audit log + no double charges
Card providers: Lithic, https://t.co/LTtoTDK4sS & Airwallex (beta) + mock cards for testing.
“Let your agent buy, safely.”
https://t.co/kttN7gdlEX
Star it, try the mock rail, or open a PR.
@Lyra0x_ Yes. Approval is tied to the exact purchase. The retry key includes merchant + amount, so changing either = a new request needing fresh approval. The card is capped at the approved amount, single-use, and each charge must match the approved merchant. good question lyra @Lyra0x_
1/ I need some honest advice.
In January I built a Xero tool off a feature request with 528 upvotes. 4 people used it and liked it. Nobody paid, because I never asked. I burned out and quit at month 4. 🧵
2/ Now I've built an AI-guided DNS incident resolution tool for web agencies & MSPs.
I know DNS well. I don't know agencies or MSPs.
I don't want to lose another 4 months on something nobody buys.
Day 3: I just got the Run It MVP working end-to-end.
A real DNS incident:
Detect → Diagnose → Approve → Fix → Verify → Resolve → Rollback
It found the DNS change, proposed the fix, applied it, waited for propagation, independently verified recovery, and successfully rolled back.
Not just monitoring anymore
A client site returns Cloudflare 523.
The agency now checks:
Cloudflare → origin host → DNS → SSL → WP → “did anyone change nameservers”
That’s 6 dashboards and 40 minutes
before anyone even proposes a fix.
That’s the workflow we’re collapsing.
Writing all 10 up properly. If you've hand-installed any of these and hit something ugly, tell me which — I want the failure modes I haven't found yet.
I ran 10 popular open-source apps through my auto-deployer this week.
1 came out live.The other 9 broke in 9 completely different ways — and almost none of them were the AI getting it wrong. Every single one:
And the humbling one — I'd written off https://t.co/KEk5fzZdCW as unfixable. Then I fixed something unrelated (build logs were being discarded on failure) and the "unfixable" blocker was a missing Python in one build stage.
My agent couldn't fix what my error handling deleted.
So instead of "what would you use this for" —
What's a tool you pay for where you found the open-source version, opened the README, and closed the tab?
Reply with what it was and what stopped you. That's what I build next.
Every SaaS you pay for has a free open-source version sitting on GitHub right now.
You don't use it because getting it running is a weekend you don't have.
I made that a paste-a-link operation. This is Plane — a free Jira — going live:
Here's where I actually need help.
I know the machine works. I don't know who it's for.
Cheaper than SaaS? Data that legally can't leave your servers? An agency standing up the same stack for 20 clients? Those are three different products.