Ghostty Is Leaving GitHub: Why It's Time to Run Your AI Agent Locally

When GitHub user #1299 packs up and leaves after 18 years, you're not watching a tantrum. You're watching the canary in the coal mine.

On April 29, 2026, HashiCorp co-founder Mitchell Hashimoto announced that Ghostty, his fast, GPU-accelerated terminal emulator, would be migrating off GitHub. He had kept a journal for the previous month, marking an "X" on every day a GitHub outage blocked his work — and almost every day had one. The platform that had been his daily home since February 2008 had become, in his words, no longer a place for serious work.

This isn't drama. It's a leading indicator. Centralized developer infrastructure is buckling under the weight of AI agents — and the only durable fix is to run your AI agent locally.

GitHub's Worst Week in Recent Memory

To understand why a long-time loyalist would walk away, look at what happened in a single five-day window:

April 23, 2026 — The Silent Revert. A regression in GitHub's merge queue caused pull requests merged via squash to produce incorrect commits when a merge group contained more than one PR. The result: changes from previously merged PRs were silently reverted. The official tally was 230 repositories and 2,092 pull requests affected during the impact window. No errors. No warnings. Code just… disappeared from main branches.

April 27, 2026 — Search Collapses. The Elasticsearch backend powering GitHub Search was hammered by what GitHub described as a likely botnet, taking down search across pull requests, issues, and projects for hours.

April 28, 2026 — Critical RCE Disclosure. Wiz researchers disclosed CVE-2026-3854, a remote code execution vulnerability where a single git push with crafted options could run code on GitHub's servers. CVSS 8.7. Reports indicated 88% of GitHub Enterprise Server instances remained unpatched at disclosure.

Three failures of git's trust contract in five days. And independent monitors painted a bleak overall picture: third-party tracking showed roughly 90.21% uptime over the previous 90 days — far below the 99.9% "three nines" SLA the platform promises.

Why a Long-Time Believer Walked

Hashimoto isn't a casual user. He built Vagrant in part because he hoped it would land him a job at GitHub. He had used the platform every day, multiple times a day, for over 18 years — saying it was where he had been "historically been happiest."

When that user leaves, the structural problem isn't the user. It's the platform.

His framing in the farewell post is the line everyone is quoting: he wants to ship software, and the platform won't let him. After 18 years, he's gone. He's not the first — the Zig programming language migrated to Codeberg in December 2025, citing similar concerns.

AI Agents Are Breaking Centralized Infrastructure

Here's the part that should keep every founder, platform engineer, and indie developer up at night: GitHub knows what's happening, and they're losing the race.

In an April 28 availability post, GitHub's CTO walked through the numbers. The company began executing a plan to scale capacity 10X in October 2025. By February 2026, it concluded it actually needed to design for 30X today's scale. The driver: agentic development workflows that accelerated sharply after the second half of December 2025.

The traffic shape is brutal. GitHub COO Kyle Daigle reported the platform went from 1 billion commits in all of 2025 to 275 million per week — a pace of 14 billion per year if linear. GitHub Actions usage went from 500 million minutes per week in 2023, to 1 billion in 2025, to 2.1 billion in a single week recently.

This isn't more humans. It's fleets of AI coding agents — Claude Code, Cursor, Codex, and others — autonomously cloning repos, opening PRs, running CI, and iterating at machine speed. The platform's free-tier limits and Actions minutes were designed for humans and the occasional bot. They were never sized for swarms.

And here's the architectural punchline. A single pull request now touches Git storage, mergeability checks, branch protection, GitHub Actions, search, notifications, permissions, webhooks, APIs, background jobs, caches, and databases. At high scale, small inefficiencies compound — queues deepen, cache misses become database load, indexes fall behind, retries amplify traffic.

That's the real story behind the merge queue silently reverting your code. It's not one bug. It's coupling.

When Everything Runs in the Cloud, Their Bad Day Is Your Bad Day

This is the line that should land for anyone building serious software in 2026.

If your AI agent lives in someone else's cloud, you don't have a tool — you have a dependency. When GitHub's Elasticsearch falls over, your agent's PR list returns empty. When a merge queue regression ships, your agent's "successful" merge silently reverts a colleague's commit. When a CVE drops and the platform takes hours to patch, your supply chain takes hours to patch.

Every layer of centralization is a layer of shared fate.

The irony: Git itself was designed to be decentralized. Linus built it for Linux kernel development across thousands of contributors with no central authority. The protocol assumes the network might fail. The ecosystem we built on top of it assumed the opposite — that a single platform would always be there.

That assumption is now demonstrably wrong.

The Local-First Fix

The fix isn't a "better cloud." Switching from one centralized platform to another just changes whose bad day becomes yours. The fix is local-first AI.

Local-first means your agent runs on your machine. Your code never leaves your laptop unless you push it. Your agent doesn't queue behind 90 million monthly PRs from someone else's pipeline. Your agent doesn't get rate-limited because a botnet is hammering an Elasticsearch cluster three regions away. Your agent doesn't silently revert your colleague's work because of a regression in someone else's merge queue.

When the cloud goes down, you keep coding.

OpenSair: Your Agent. Your Mac. Your Day.

This is exactly why we built OpenSair. OpenSair is a local-first AI agent that runs entirely on your Mac. No cloud round-trip. No platform outages. No single point of failure. No swarm of other people's agents in front of yours in line.

The promise is simple:

Your agent runs on your machine. When AWS, GitHub, or anyone else has a bad day, you don't.
Your code stays local by default. Privacy isn't a setting — it's the architecture.
Your throughput isn't shared. You're not competing with global agentic traffic for compute or API quota.
Your tooling is yours. No vendor lock-in. No surprise pricing migrations. No "usage-based" billing changes mid-quarter. We're not betting against the cloud. We're betting against single points of failure — and in 2026, that's a bet the GitHub status page is making for us every week.

The Bigger Picture

Hashimoto leaving GitHub isn't just a story about one terminal emulator. It's a structural signal. The platforms developers built their workflows on were sized for a pre-agentic world. The agentic world is already here, and the platforms are visibly straining.

You have three options:

Hope your provider scales 30X faster than the world generates AI traffic. Good luck.
Move to another centralized provider and hope they get it right. Same bet, different logo.
Pull your agent local. Own your tools. Own your uptime. Own your day. The first two ask you to trust someone else's roadmap. Only the third puts you in control.

Run Your AI Agent Locally

Ghostty is the canary. The merge queue bug is the warning. GitHub's own 30X capacity admission is the alarm.

The fix isn't a better cloud. It's local-first AI.

OpenSair runs entirely on your Mac. No cloud. No outages. No single point of failure.

Your agent. Your machine.

👉 Get OpenSair at opensair.ai

Sources: Mitchell Hashimoto's blog post "Ghostty Is Leaving GitHub" (mitchellh.com), GitHub Engineering Blog "An Update on GitHub Availability" (April 28, 2026), GitHub Community Discussion #193645 (April 23, 2026 incident thread), Wiz security disclosure for CVE-2026-3854, The Register, Linuxiac, XDA Developers, OfficeChai.