Key takeaways
- Archived nine days after launch. The final commit landed March 24, 2026, and the README now opens with "This project is no longer actively maintained. Fork it, customize it, make it yours"
- The creator's post-mortem is the real artifact — "the terminal already won," "delegation is a prompt, not a product." A rare honest writeup of why custom multi-agent GUIs lose to the native TUIs
- The code works and remains MIT-licensed — multi-agent grid, coordinator/worker delegation, SSH remote targeting, Gemini CLI support, and 93 passing tests live on the legacy branch at commit 78f1bf2
- Traction outlived the project — 1,363 stars and 83 forks as of June 2026, up from 243 at launch, driven by an X announcement that opened "Karpathy asked. I delivered"
FAQ
What is OpenSquirrel?
A native macOS app (Rust/GPUI/Metal) for running multiple AI coding agents side by side — Claude Code, Codex, Cursor Agent, Gemini CLI, and OpenCode — with coordinator/worker delegation, SSH remote targeting, MCP integration, and persistent sessions. It was archived by its creator on March 24, 2026, nine days after launch, but the code remains buildable under MIT.
Why was OpenSquirrel archived?
Creator Elliot Arledge concluded "you don't need a custom GUI to orchestrate AI agents." His post-mortem argues every agent CLI already ships a polished TUI, delegation can be done with a few lines in a CLAUDE.md file, and building a rigid UI around an unsettled workflow is premature. He now recommends a plain terminal plus config-file delegation instead.
How does it compare to Okena?
Both use Rust/GPUI. Okena focuses on terminal multiplexing with cross-platform support and is still maintained. OpenSquirrel focused on multi-agent orchestration with coordinator/worker delegation and remote SSH targeting, but was abandoned in March 2026 — its author's own conclusion was that the terminal-native approach wins.
Overview
Status: archived. OpenSquirrel was abandoned by its creator on March 24, 2026 — nine days after launch. The README now opens: "This project is no longer actively maintained. Fork it, customize it, make it yours." The repository is not GitHub-archived and remains buildable, but the final commit ("Archive project: update README with learnings") landed March 24, 2026.
OpenSquirrel was a native, GPU-rendered control plane for AI coding agents built in Rust on GPUI (the UI framework from Zed). It ran Claude Code, Codex, Cursor Agent, Gemini CLI, and OpenCode side by side in a responsive tiled layout — agents auto-arranged based on count (1=full, 2=split, 4=2x2, etc).
The tagline said it all: "For people who get distracted by agents." Instead of switching between terminal tabs, OpenSquirrel gave you a single view of all your agents working simultaneously. It launched with a splash — creator Elliot Arledge's announcement opened "Karpathy asked. I delivered."
Key stats (as of June 11, 2026): 1,363 stars, 83 forks, MIT license, Rust, macOS (Metal) with Linux (Vulkan) compiling. Created March 15, 2026; last commit March 24, 2026.
Free and open source — no pricing, no funding; this was a solo project, and no funding was ever publicly disclosed.
Architecture (as archived)
The full feature set lives on the legacy branch at commit 78f1bf2.
Multi-Agent Grid
Multiple agents ran simultaneously in a GPU-rendered tiled layout. Each agent got its own pane with structured output parsing — custom markdown rendering with code blocks, diffs, headings, and bullets, all parsed from stream-json output across runtimes.
Coordinator/Worker Delegation
The standout feature: a primary agent (e.g., Opus) could automatically spawn sub-agents for focused tasks. Workers executed and returned condensed results, not full transcripts — hierarchical task decomposition across heterogeneous agents (Claude Code + Codex + Cursor simultaneously). The author's post-mortem later disowned this as a product: the same behavior, he found, takes about four lines in a CLAUDE.md file telling the model to invoke external CLIs headlessly via Bash.
Remote Machine Targeting
Agents could target local or remote machines via SSH + tmux, configured in ~/.osq/config.toml, with agents assigned to specific hosts — Opus coordinating locally while workers executed on remote servers.
Supported Runtimes
| Runtime | Mode |
|---|---|
| Claude Code | Persistent stdin (multi-turn) |
| Codex | One-shot per prompt |
| Cursor Agent | One-shot per prompt |
| Gemini CLI | One-shot per prompt |
| OpenCode | One-shot per prompt |
All runtimes ran with permission-bypass flags (e.g., --dangerously-skip-permissions, --yolo) — a security tradeoff inherent to the design.
Features Added During Its Nine-Day Life
The March 15–24 commit history shows a fast burn: a "superset rewrite" with a daemon, settings, and Linux support (March 18); 26 security and edge-case tests from an audit (March 18); and a UI overhaul with reusable components, a fuzzy model picker (Cmd+M), ephemeral workers, and per-agent token/cost tracking (March 23). The archived build ships MCP integration, 7 themes, persistent sessions, and 93 passing tests.
Why It Was Archived
The README post-mortem is unusually candid and worth reading in full. The core conclusion: "you don't need a custom GUI to orchestrate AI agents." The arguments:
- The terminal already won. Every agent CLI ships a polished TUI; a Rust GUI re-rendering their JSON output "will always be worse than just... using the native TUI."
- Delegation is a prompt, not a product. No orchestration daemon, hooks, or middleware needed — just CLAUDE.md instructions.
- The architectural mismatch. Using Claude Code to iterate on a GPUI render pipeline meant "every feature took 10x longer than it should have."
- Models aren't good enough yet for opinionated UX. Multi-agent workflows haven't settled; rigid UIs around one workflow lock you in.
His replacement stack: Ghostty, agents run directly, delegation via ~/.claude/CLAUDE.md, token tracking via CodexBar, multi-agent via terminal tabs.
Viability
None as a product — this is an archived solo project. The aliveness signals are unambiguous: zero commits since March 24, 2026, an explicit abandonment notice in the README, no releases or tags ever cut, and only 4 open issues against 8 watchers as of June 11, 2026. The 1,363 stars reflect launch-window virality, not an ongoing community. There is no maintainer, no roadmap, and the creator has publicly argued against the product category itself.
What survives is the code (MIT, buildable, 93 passing tests) and the post-mortem, which has become the project's real contribution to the Mac coding-agent-app conversation.
Cautions
- Explicitly unmaintained — no commits since March 24, 2026, and the README tells you to fork it rather than expect updates. Agent CLIs change their stream-json output formats frequently; the parsers will rot.
- The author's own verdict is the strongest criticism: "Building a rigid UI around one workflow locks you in. The terminal is infinitely flexible."
- Permission bypass by design — every runtime ran with sandbox/approval bypass flags; running the archived build today carries those same risks against newer agent versions.
- Manual cargo build only — no releases were ever published; you need Rust 1.85+ and a Metal-capable Mac.
What Developers Say
No substantive third-party discussion exists — searches of Hacker News (Algolia) and Reddit as of June 11, 2026 surface no threads about OpenSquirrel. The quotable record is the creator's own arc, from launch hype to post-mortem:
"Karpathy asked. I delivered. Introducing OpenSquirrel! Written in pure rust with GPUI (same as zed) but with agents as central unit rather than files." — Elliot Arledge, launch announcement, March 2026
"The terminal already won. ... Building a Rust GUI that parses their JSON output and re-renders it will always be worse than just... using the native TUI." — Elliot Arledge, archive post-mortem
"Delegation is a prompt, not a product." — Elliot Arledge, archive post-mortem
Bottom Line
Not recommended as a tool — it's archived, unmaintained, and its agent-output parsers will drift out of sync with the fast-moving CLIs it wraps. Recommended reading, though: the post-mortem is one of the most honest documents in the Mac coding-agent-app category, and its thesis — that orchestration is a prompt-and-terminal problem, not a GUI problem — is a direct challenge to every other app in this comparison. Outlook: the repo will likely persist as a reference implementation and cautionary tale; any future life would come from a fork, of which none prominent exist as of June 2026.
Research by Ry Walker Research