A Self-Improving Pipeline for $110 a Month

A developer built a system called autoloop to automate the manual work of handling GitHub issues. It combines Claude Max with a simple VPS, running at a low cost of about $110 a month (Claude Max $100 + VPS $9.99).

How the system works

Three systemd timers run automatically at fixed times (midnight, 2 AM, 4 AM). GitHub issues go through the following stages:

• Triage (Sonnet, ~$0.10): Reads and validates the issue and estimates story points. Issues that are too large are automatically broken down into smaller sub-issues, taking dependency order into account.
• Implement (Opus, ~$1.50): Selects issues starting from the top of the dependency chain, creates a branch, has Claude write the code, runs tests, and automatically opens a PR.
• Verification: If tests fail or test files are missing, it retries up to 3 times; if all retries fail, it attaches a needs-human label and moves on to the next item.

The developer's only role is reviewing and merging PRs from their phone.

Results over two weeks

27 PRs merged, 1 failure, with an average cost of $1.61 per issue. The system's failures were mostly cases where the issue description was unclear or lacked sufficient context. The developer noted that "good issue specs make good PRs."

Safety design

The core safety measure is that the system cannot modify itself. The autoloop/ and autoloop.toml paths are protected, and any issue targeting these paths is automatically routed to needs-human. Self-improvement is possible, but self-modification is not.

Scope and limitations

This pattern suits solo developers, small teams, and non-critical systems where a 24-hour fix cycle is acceptable. It is not yet suitable for regulated environments, multi-team codebases, or customer-facing production systems that require rollback mechanisms. The author noted that two weeks of data isn't long-term validation, and promised further reports after 3 months and 6 months.