Agent PromptingTurn a vague ask into a job an AI agent can actually run
MCPlato interviews you, writes the prompt contract, and saves a reusable agent brief.
Sam wanted Claude Code's loop without leaving his machine. MCPlato reads the repo, reproduces the failure, patches it, re-runs the suite until it is green, and ships the PR — locally, with his approval.
Meet
Sam Okafor
Indie Developer · Solo maintainer · 1 OSS CLI + 2 side projects
Sam ships alone. Between flaky tests, stale issues, and changelogs nobody writes, the boring half of maintaining three repos eats the time he wants to spend building. He wants an agent that works in the codebase, not just a chat window.
Prompts to try
The auth token-refresh test is flaky and it's blocking my release. Find the real bug, fix it, and prove it's fixed. - Run the suite and reproduce the actual failure — don't guess - Read src/auth/refresh.ts and trace the root cause of the flakiness - Patch it, then re-run the tests and keep correcting until everything is green - Open a PR with the diff, the test results, and a short summary of what was wrong Work in my local repo and don't push to main — just open the PR for review.
Describe the bug or the task in the repo. MCPlato works directly in your local working tree — your files, your toolchain, your environment.
It runs the suite, watches it fail, and reads the relevant source to find the root cause — not a guess from a snippet, the actual failing run.
It patches the code, runs the tests again, and keeps correcting itself until the suite passes — the self-improving loop, on your codebase.
Open a PR with the diff, test results, and a summary — or just commit to a branch. Schedule a daily CI digest so nothing rots.
Local
agent on his real codebase
Red → green
reproduce, fix, verify — no babysitting
3 repos
maintained without the busywork
MCPlato is a desktop app, not a terminal-only tool. You get the same read-fix-test-repeat loop on your local codebase, plus scheduled tasks, MCP integrations, a built-in browser, and approval modes — in one workspace.
Yes. It works in your local working tree with your real toolchain — it runs your test command, reads the failing output, and iterates against that, not against a copied snippet.
No. It stops at a checkpoint and lets you choose: open a PR, commit to a branch, or wait. With an approval mode, nothing leaves your machine until you say so.
Yes, via MCP. It can open PRs, read issues, and comment, so the whole reproduce-to-PR flow stays in one place.
Schedule it to run a CI digest, triage new issues, or sweep for flaky tests across each repo, and it reports back so the busywork never piles up.
Reproduce, fix, verify, ship — locally, with your approval.