September 21, 2026
An agent now handles large PRs in the cloud: Capy fixes failed CI and pushes again
Capy manages the branch itself: CI fails, the agent reads the error, fixes it, and pushes again. There is no local installation, and the first week costs $1.
Garry Tan
@garrytan
I do not understand how @capydotai does it, but it really handles multi-step tasks and ships large PRs faster than Codex or Claude Code, entirely on its own. My favorite new secret weapon for agentic coding over the past week.
· 89.8K views
The agent works in a separate capy/ branch and does not touch main. If CI fails, it wakes up on its own and pushes a fix.
Long-running tasks usually break right here: checks go red, and a person has to return to steer the agent manually. Capy puts dependent tasks into a stack, with each next PR built on the previous branch, so there is no need to wait for a merge.
Nothing to install. Codex and Claude Code have a CLI on your own machine; Capy has none at all. The flow is: go to capy.ai, install the GitHub App, choose repositories, choose a model, and send the first prompt.
Billing runs on credits. The first 7 days cost $1 with a verified card; after that, the Lite plan is $20 per month, and those same $20 are added as credits. They pay for model tokens and virtual-machine time, with the machine priced at $0.40 per hour by default. Any unused balance expires at the end of the month.
You do not have to pay Capy for tokens: some of its 40-plus models run through your Codex, Copilot, SuperGrok, and Azure subscriptions and are charged to them. The connection is managed in model settings.
The agent can also run without a person: on a schedule (at least once every 5 minutes), on GitHub events, from a Slack message, or through a webhook.
Behind Capy is San Francisco-based Scrapybara: the vendor claims an audience of 70 thousand engineers.
Source
