Build or use · 2026
Build your own AI developer or use mcptask?
A basic “task → Claude Code → pull request” loop is a weekend script, and GitHub Copilot and Codex offer one too. Many software companies already run something like it. The hard part is everything around the loop that keeps it working every day without a person watching. Here is that list, row by row.
What a production runner needs beyond “task → PR”
What a typical internal script has, next to what mcptask.online does today.
| What it takes | Typical internal script | mcptask.online |
|---|---|---|
| Watchdog and crash recovery | Usually none: a hung CLI waits until someone notices | A watchdog ends a frozen run, a stall detector catches a stuck one, and a run that crashes or overflows its context resumes from its branch |
| Triage before each task | One fixed model for every task | Each task is triaged first and gets the model tier it needs |
| Daily budget | The spend shows up on the provider's bill afterwards | A daily budget read live during the run; the runner stops when it is spent |
| A real task queue | A label or a list the script reads in order | Priorities, bugs first, then the oldest; a blocked task waits for its blocker links |
| Time logging | None | Every run logs its progress and time against the task, ready for work reports and invoicing |
| Live dashboard | A log file on one machine | The Runner sessions panel shows each runner's state (waiting, triage, processing) as it happens |
| A fleet of runners across machines | One machine, one checkout | A runner on every machine you add, all taking work from one queue, one runner per checkout. They run on your machines, so your data stays with you |
| Coding CLIs and models | Wired to one CLI | Claude Code, Codex CLI or OpenCode, with hosted or local open-weights models |
| Hard failure | A red line in a log nobody reads | A hard failure files a bug task with the run log attached |
| Self-update | Someone pulls the new version by hand | The runner updates itself and verifies each release against its published checksums |
| Git hosts | Usually the one host the company uses | GitHub, Bitbucket Cloud and GitLab; every change is a pull request through your CI |
The part nobody budgets for: upkeep
Models and coding CLIs change every few weeks: new flags, new output formats, new limits, new models worth switching to. An internal runner is a product of its own, and someone on your team keeps it working instead of shipping your work. With mcptask.online that upkeep is ours.
The honest answer
If you already have all of this and enjoy maintaining it, you don't need us.