Changelog

Every build we have shipped, straight from GitHub Releases. Newest first.

v0.1.8 upgrades the Pi engine and adds Radius as a provider.

What changed

  • Pi upgraded to 0.80.7. The engine update adds support for Fable 5 thinking levels, a toolChoice option for providers, dynamic tool loading, and a batch of upstream fixes: fallback transcript attachments no longer stall, provider transports handle aborted streams cleanly, and several model metadata fixes ship. The RPC surface is unchanged; no frontend or sidecar protocol adjustments were needed.
  • Radius is now a supported provider. You can sign in with Radius via OAuth, or bring your own API key under the PI_GATEWAY_API_KEY env var. Radius appears in the settings panel alongside Anthropic, OpenAI Codex, and GitHub Copilot. A label-derived monogram stands in as the provider icon until an approved glyph ships.

An installed Hoy no longer mixes its Rust bridge with a stale development sidecar left over from building in the repo.

What changed

  • Installed app stays on its own sidecar. When you local:install Hoy from a repo that also has dev sidecar artifacts, the installed app now prefers the sidecar it shipped with instead of picking up whatever was last built under packages/sidecar/. This keeps the Rust bridge and sidecar protocol in sync after a self-update or rebuild.

v0.1.6 upgrades the Pi engine and fixes MCP server reliability when Hoy is launched from a desktop environment.

What changed

  • Pi upgraded to 0.80.6. The engine update adds a new "Max" thinking level above X-High, switches to agent_settled as the proper terminal boundary for turn completion (so post-run compaction and queued continuation events now reach you), and inherits a batch of upstream fixes: truncated tool calls no longer hang, Windows project context traversal is fixed, retry classification is sharper, and stale pre-compaction usage no longer skews budget estimates.
  • MCP servers now see your real PATH, not the desktop launcher's minimal one. Hoy resolves your full login-shell PATH before spawning sidecar processes. Previously, MCP server configs referencing ${PATH} overwrote that resolved PATH, breaking commands like bunx. Now fixed.
  • Login shell resolution is more robust. On NixOS and Guix systems, Hoy queries the system passwd database to find your login shell (store-path aware) instead of relying only on $SHELL. Shell candidates are deduplicated. A thread and child process leak on shell timeout is fixed, and chatty shell init scripts no longer deadlock the PATH query.
  • Windows MCP commands resolve correctly. resolveCommand() now iterates over PATHEXT extensions (.cmd, .bat, .exe, .com) so npm-installed CLI shims like bunx.cmd are found without the MCP transport ever seeing the bare command name.

Subagent threads stay out of your home recents, MCP servers launch reliably, and FleetView is now Fleet.

What changed

  • Subagent conversations no longer appear as recents on the home page. Only the conversations you start yourself show up there. Spawned subagent threads are still visible in Fleet.
  • FleetView is now Fleet. The subagent dashboard's header is simpler (the back-arrow button is gone; just the title sits at the top) and it gets a shorter name. This is a rename and a cleanup, no behavior change.
  • MCP servers resolve their commands against your full PATH. Relative commands like bunx or uvx now work without an absolute path. And when Hoy launches from a desktop shortcut that strips your shell environment (AppImage, .desktop file), it still picks up the PATH entries from your login shell so tools managed by version managers (nvm, fnm, volta, bun) are found.

A system tray icon so Hoy stays one click away, plus refreshed app icons across every platform.

What changed

  • A tray icon with Show/Quit. Hoy now lives in your system tray. Click the icon to bring the window to focus, or right-click for Show Hoy and Quit. The window unminimizes and grabs focus even if it was hidden.
  • Refreshed app icons. All platform icons (macOS, Windows, Linux) have been regenerated at higher resolution for sharper display on retina and high-DPI screens.

An onboarding flow for first-time setup, theme support, and fixes that make Hoy launch correctly from desktop shortcuts.

What's in it

  • An onboarding flow for your first launch. Hoy now guides you through setup the first time it runs: pick a theme, connect one model provider, and you are in. Already have a key? No flash of chrome while it checks: just a quiet pulse until the provider status comes back, then straight to work.
  • Light, dark, and system themes. The Appearance panel in settings is no longer a placeholder. Pick dark (the layered near-black workbench), light, or follow your OS setting. The default on fresh installs is system, so Hoy respects your OS preference out of the box.
  • Three more model providers. The provider list now includes Ant Ling, NVIDIA NIM, and ZAI Coding Plan (China), matching every provider Pi 0.80.3 supports. The ZAI entry is now labeled "ZAI Coding Plan (Global)" for clarity, and the moonshotai-cn env var name is corrected.
  • Hoy finds your shell tools. When launched from a .desktop shortcut (which strips most of PATH), Hoy now discovers your real PATH from your login shell. Version-manager tools like bun, node, or uv are visible to the agent.
  • Debug builds stay out of production's way. The dev instance now uses its own DBus ID, so a running production Hoy no longer blocks the debug build from opening, and vice versa.
  • Sidebar state survives a restart. Collapse the sidebar, close Hoy, reopen it: it stays collapsed.

Fixed

  • No more full-chrome flash during startup. While Hoy checks provider status at launch, it now shows a steady logo pulse instead of briefly rendering the full app layout and then switching to onboarding. Onboarded users with a saved key should never see the flash again.

The first big feature release since the initial build: your agent can now spawn a whole fleet of agents, plan before it acts, work toward a goal on its own, and branch a thread wherever you want to explore.

What's in it

  • Agents that spawn agents, on one screen. Your agent can now delegate to subagents, which can delegate further, all running at once under a concurrency cap. FleetView gives you a live tree of every agent in the run, what it is doing right now, and its rolling token and cost totals. Results flow back into the parent thread inline the moment a child finishes. You can author your own subagent types (create, edit, duplicate, delete) with their own model, prompt, and tool scope.
  • Plan Mode. Ask for a plan and Hoy switches into a two-phase flow: it clarifies the goal and architecture with you first, then proposes a task-by-task plan as an interactive card. Approve it and the tasks execute one by one, each handed to a subagent, with per-task oversight available if you want to review every edit.
  • Goal Mode. Give the agent a goal with /goal and it keeps working toward it across turns, checking its own progress. Optionally gate each round on a verify command (tests, a build, a lint) or a read-only auditor pass, so it only counts a goal as met when the check actually passes.
  • A session tree you can branch. The new /tree navigator opens in a right-side dock and maps every entry in a thread. Click any node to jump the transcript to it, or branch from that point to fork a new thread, or clone the whole thread to explore an alternative without losing the original.
  • A home that starts you working, plus a usage dashboard. The home screen is now a start-a-task hero that opens straight into a new thread on your last project. A dedicated Usage view charts your token trends over time, breaks usage down by model, and shows an activity heatmap and streaks.
  • MCP servers. Connect Model Context Protocol servers to give your agent more tools; add one by pasting its JSON config in settings.
  • Skills. Invoke a skill from the composer to run a packaged, reusable workflow in the current thread.
  • A richer composer. An @ context picker with inline mention chips pulls in files, threads, images, and commands; slash commands autocomplete as you type. Steer the agent mid-turn or queue follow-up messages while it works, and answer structured clarifying questions one step at a time when the agent needs a decision.
  • Faster and calmer to watch. Reopened threads render instantly from disk before the agent process is even ready, streaming output is coalesced to one repaint per frame so long turns stay smooth, a live timer shows thinking as it happens, and tool calls collapse by default so the transcript stays readable.
  • Quality-of-life throughout. Auto-compaction keeps long threads within context (with a manual compact when you want it), the machine stays awake while an agent is streaming, a second launch focuses the running window instead of opening another, threads carry their own identity color and model glyph in the sidebar, and you can sign in to a provider with a subscription via OAuth.

The first public build of Hoy: a real desktop app that runs your coding agent on your own machine, on your own keys. It's beta, so expect rough edges.

What's in it

  • Your agent, in a real window. Prompt, watch it work, and steer it mid-run from a desktop app instead of a terminal. Every file it reads, every edit, every command lands in the thread the moment it happens, result attached.
  • Threads that stick around. Sessions are saved to your disk and come back exactly where you left them after a restart. Nothing lives on a server we run.
  • Your keys, every model. Drop in a key for any provider Pi supports, Anthropic, OpenAI, Google, Azure, Bedrock, Mistral, Groq, xAI, OpenRouter, Ollama, and more, and switch between them from one menu, mid-conversation. Keys are written to a local, private credential file (mode 0600) and never leave your machine.
  • Local, end to end. Hoy drives the Pi agent as a separate process on your machine. No cloud account, no sign-up, no middleman.

Install

Builds for macOS (Intel and Apple Silicon), Windows, and Linux (.deb, .rpm, .AppImage) are attached below.

The direct downloads aren't signed yet, so your OS will warn you the first time; click through once and you're in. On Arch, the pacman repo is signed end to end.

Known rough edges

  • Pre-1.0 and experimental; expect bugs.
  • Direct downloads are unsigned (the Arch pacman repo is the exception).
  • Bring your own API key; there's no hosted option.