You file it. Buttons appear.
Pullfrog posts Quick Links on new issues: Make a plan, Implement. Or tag @pullfrog in the body to start immediately.
Pullfrog · now free for personal + open source
Pullfrog is an open-source, model-agnostic GitHub bot from Zod creator Colin McDonnell. It listens for repo events — issues opened, PRs created, reviews submitted, CI failures — and runs coding agents inside your own GitHub Actions. On August 8 it went free for personal accounts and open source.
Public beta announced May 12, 2026; now generally available. Bring your own keys or use Pullfrog Router at raw provider cost with no markup — either way, compute runs on your Actions minutes and your model spend.
new interfaces to learn. Every step above happens as comments, branches, and checks in the repo you already have open.
The pond is the event loop
Every hop is a GitHub webhook event; every landing is an agent run in your own Actions. Scrub through a full plan → implement → review → fix cycle.
Pullfrog posts Quick Links on new issues: Make a plan, Implement. Or tag @pullfrog in the body to start immediately.
Iterate by replying — or just edit the plan comment directly yourself. When it reads right, hit Implement plan.
The agent runs against the latest version of the plan and opens a PR linked back to the issue.
It addresses comments and pushes updates. Batch with Fix all, or 👍 the comments you agree with and hit Fix 👍s.
With guardrails: skips bot-authored commits, only runs when the target branch is green, and rate-limits retries.
Git and GitHub operations go through a permission layer — the agent can't push to protected branches.
Install the GitHub App, add one workflow file — pullfrog.yml — and Pullfrog can dispatch agent runs in response to webhooks. Keys live in GitHub's secret store, get auto-masked, and each run uses a short-lived installation token that's revoked when the run completes. Works with Anthropic, OpenAI, Google, Mistral, DeepSeek, OpenRouter — or an existing Claude Code or Codex subscription.
BYOK to any provider, or use Pullfrog Router — billed at raw provider cost with no markup — and skip key management entirely.
McDonnell describes Pullfrog as a harness over OpenCode and Claude Code intended to run in CI. For local development, he says, just use those agents directly.
shell in an isolated subprocess without sensitive env vars; git and GitHub ops through a purpose-built MCP server with permission checks; a built-in headless browser for E2E tests and screenshots.
on pullfrog/pullfrog — early numbers for a GA tool. The landing page's "trusted by" row lists Remotion, Zod, and ArkType; Zod is McDonnell's own project.
When someone reviews a PR Pullfrog created, it picks up the feedback and pushes updates on its own. When a check suite fails, it can read the logs and attempt a fix — on its own PRs by default, and optionally on human PRs it has reviewed.
Auto-fix skips bot-authored commits, only runs when CI is already passing on the target branch, and rate-limits retries to avoid runaway fix loops.
Batch-address a whole review, or react 👍 to the comments you agree with and Pullfrog applies only those, resolving the threads as it goes.
The August 8 pricing change dropped per-seat math entirely: personal accounts and public repos cost nothing, and an organization running Pullfrog on private repos pays one flat monthly fee no matter how many seats or repos.
An Enterprise tier is listed as coming soon. "Free" still meters elsewhere: agent runs consume your GitHub Actions minutes and your model tokens (or Router balance) either way.
The agent didn't get a new interface. Your repo became one.
@pullfrogmake a plan, then fix it