Claude tried to earn $20 on open-source bounties. It made $0.

Trying to earn money with Claude on open-source bounties was the premise of a small, honest experiment published this week. Developer ztc00 built algora-scout, pointed Claude at it with a $20 token budget, and walked away with nothing. This post covers what the slug 2026-05-17-i-tried-to-make-claude-make-me-money-on-open-sourc refers to, and why the zero matters more than a win would have.

The result is more useful than a success story would have been. If you were wondering whether you can plug Claude into a bounty board and let it pay your AWS bill — this is the post that saves you the $20.

---

🔍 What the experiment actually did

The author wrote a small Python tool, scout.py, that searches GitHub for issues tagged with a Bounty label. The scout scores each issue on several signals before deciding whether to attempt a fix.

Signal	What it checks
Dollar value	Filter out sub-$5 spam-tier bounties
Prior attempts	How many agents have already tried this issue
Assignees	Is anyone officially working on it?
Open competing PRs	Is it already being solved in parallel?
Staleness	Assigned more than 14 days ago with no PR = potentially ripe

Claude acted as the worker with access to gh CLI, Git, and Bash. A human reviewed every PR before submission. The system scanned roughly 60–80 fresh bounties across multiple sessions.

Here is what came back:

Category	Count	Outcome
Spam-tier (less than $5)	~30 issues	Not worth the round trip
Legitimate ($50–$1,000)	Most of the rest	8–158 attempts logged within hours
"Claimed but stale" ripe targets	~0 viable	Zero candidates across multiple scans
Net earnings	—	$0

---

📊 Why that zero is the real signal

The viral tweet that inspired the experiment claimed roughly $16.88 in earnings against roughly $16 in token spend. That is not a business model. That is breaking even on coffee money before you count the time a human spent reviewing every PR.

The actual finding: Public Algora boards are now agent-saturated. By the time a human reads the issue description, twenty or more attempts have already been logged. Being the eleventh PR in a queue is not a coding problem — it is a queuing problem.

If a tool is easy to point at a public marketplace, every coding agent on the planet is already pointed there. The marginal value of running yours too is approximately zero.

The author's own conclusion was blunt: skip public boards, look at private vulnerability platforms like HackerOne or Bugcrowd instead, and build real maintainer relationships before expecting to earn anything.

---

💰 What this looks like on a Sri Lankan budget

$20 in Anthropic credit converts to somewhere above Rs 6,000 at current rates — a month of mobile data, or three weeks of lunch. It is not nothing.

For a self-funded developer or a university student in Sri Lanka, the opportunity cost of burning that credit on a saturated market is real. The question worth asking is not "can I beat the agent farms?" It is "where does my $20 produce a return I can compound?"

Use of $20 in Claude credit	Expected return
Public bounty boards (e.g. Algora)	$0 — market is saturated
Build a tool you would otherwise pay SaaS for	Time saved indefinitely
Write tests or docs for a project you already use	Maintainer trust + contribution history
Explore private bug bounty platforms (HackerOne)	Real money, but needs a track record first
Automate a repetitive local-client workflow	Direct billable value to an existing client

For a developer at a Sri Lankan software house earning in LKR, a $20 spend that returns nothing stings more than it does for someone billing in USD. The asymmetry matters when deciding where to point your token budget.

---

🧪 The deeper lesson about AI agent economics

There is a pattern here that goes beyond bounty boards. Whenever an AI-assisted workflow becomes easy enough for anyone to run, the competitive advantage disappears fast. The half-life of a "Claude makes me money" hack is measured in weeks, not months.

The moment a technique is bloggable, it is probably already too late to profit from it. The window exists between "this works" and "everyone knows this works."

What still has value is using AI in places where the output is hard to commoditise: client-specific context, domain knowledge that is not on GitHub, or problems with a small enough audience that no agent farm has bothered to build a scanner for it.

Approach	Commoditisation risk
Public bounty board with generic agent	Very high — reproducible by anyone
Automated vulnerability scanning for a niche stack	Medium — requires domain knowledge
AI-assisted billing or workflow automation for local clients	Low — context is unique to each client
Building tools for underserved local markets	Very low — no one else is building for that audience

The last row is the one this site is built on. Tools targeting Sri Lankan tax brackets, EPF rules, and public holidays are not being built by a San Francisco startup. That specificity is a durable advantage.

---

💡 What this means for you

Spend your Claude tokens on work where you — not the maintainer queue — decide whether the output has value.

The public bounty board has become a coin-pusher arcade run by other people's agents, and the house always wins. The algora-scout post is worth reading in full: honest devloop, honest budget, honest zero. Redirect the $20 accordingly.

If you want to calculate what freelance or foreign-platform earnings actually net after tax and fees, the tools below handle that math.

🔗 Useful Tools

• [Sri Lanka [Income Tax](https://induwara.lk/tools/tax-calculator-global) Calculator](https://induwara.lk/tools/sri-lanka-tax-calculator) — work out what freelance income means after APIT deductions
• Freelancer Hourly Rate Calculator — price your time before spending tokens chasing someone else's bounty
• AI Token Counter — estimate your Claude, GPT, or Gemini spend before committing a budget