Open-source developer toolBuilt from scratch, January 2026

Agent Tank: a developer tool built from an empty repository.

Agent Tank is a local dashboard and API that shows how much of your Claude Code, Codex and Antigravity subscription limits you have used. The ProPR team built it with ProPR: 46 of its 56 merged pull requests were written by agents and merged after review.

Public record

The agenttank.io home page: the headline All your coding agent usage in one place, an npm install command, and the notes Runs locally, Credentials go only to your providers, Open source (MIT).
agenttank.io, captured at 1440 pixels wide on 11 November 2026.

What it is

Agent Tank is open source under the MIT license. It runs on your own machine, reads the usage limits of the agent CLIs you already use, and shows them in one dashboard and one local API. ProPR itself uses Agent Tank to read agent capacity, so unattended runs pause before a subscription runs dry. How ProPR uses Agent Tank.

It is a first-party project: the same developer builds ProPR and Agent Tank, and the account below comes from the public record.

Where it started

The repository was created on 20 January 2026 with nothing in it, and the first agent-written pull request arrived the next day, 21 January 2026. Most of the tool landed in a single month: 42 pull requests merged in March 2026.

Its website is a separate project built the same way. That repository is private, so the counts are all that is shown here: 36 merged pull requests, 35 of them written by agents.

How the work ran through ProPR

The work followed the same short path throughout. If you have never used GitHub, three words carry the process, and the glossary has the rest:

  • Issue. A written request in plain language. "Show how fast a limit is being used up" is an issue.
  • Branch. A separate line of work where the change is made, leaving the live tool untouched in the meantime.
  • Pull request. The finished change, laid out line by line for a decision.

In ProPR, you write the issue, an agent takes it into its own isolated workspace, and a pull request comes back. Ask for a review and a second model posts numbered findings on it. You read the result, ask for changes in a comment if you want them, and merge when it is right.

Larger features used a plan: one written description, split by ProPR into several issues that each became a pull request. The "Epic" entries in the repository are those plans, with the smaller pull requests grouped under them.

Because each pull request carries one idea, the 56 of them read as a changelog of the tool, and a merged one that turns out wrong has a Revert button in GitHub.

The numbers, as of 11 November 2026.

Counts from the public repository on GitHub and from the website's private repository.

56pull requests merged
46of them written by agents through ProPR
42merged in March 2026 alone
36pull requests on its website, 35 by agents

Check the merged pull requests yourself

Example work

Five merged pull requests from the public repository, written by agents from issues. The titles are the agents' own.

What this shows for a tool of your own

A small developer tool is a good first project. A feature is an issue, the issue becomes a pull request you can read in a few minutes, and the tool keeps working the whole time because unfinished work lives on its branch until you merge it.

When a provider changes something, as happened with new models and quota pools here, the fix is another issue and another pull request. The history of the project is the list of pull requests, and anyone can read it.

To run it on your own server, self-host ProPR, free and open source.