Crewyard 0.3.0

Install

These builds are not signed with a paid certificate yet, so each system warns once before the first start, and they do not update themselves: get the next version from this page. To check a download, compare the output of Get-FileHash <file> (PowerShell), shasum -a 256 <file> (macOS) or sha256sum <file> (Linux) with its line in SHA256SUMS.txt.

Windows 11, x64: run Crewyard-Setup-<version>-x64.exe. It installs for your user only, without an admin prompt. SmartScreen says "Windows protected your PC" because the publisher is unknown: choose More info, then Run anyway.

macOS 13 (Ventura) or later: -arm64 is for Apple silicon (M1 and later), -x64 for Intel Macs. Open the .dmg and drag Crewyard to Applications, or unzip the -mac.zip. The app is not notarized, so the first start is blocked with a message that Apple could not verify it. Open System Settings, then Privacy & Security, scroll down, choose Open Anyway next to Crewyard and confirm. Or, in Terminal: xattr -dr com.apple.quarantine /Applications/Crewyard.app.

Linux, x64 (glibc 2.35 or newer: Ubuntu 22.04, Debian 12, Fedora 36 and later):

• Debian and Ubuntu: sudo apt install ./crewyard_<version>_amd64.deb. It adds Crewyard to the app menu and crewyard to your PATH, and on Ubuntu 24.04 and later installs the AppArmor profile Chromium's sandbox needs. Prefer it there.
• Any distribution: chmod +x Crewyard-<version>-x86_64.AppImage, then run it. It needs FUSE 2 (sudo apt install libfuse2, or libfuse2t64 on Ubuntu 24.04). Where user namespaces are restricted (Ubuntu 24.04 and later) its launcher starts the app without Chromium's sandbox.

API keys are kept by Windows DPAPI, the macOS Keychain, or the Linux Secret Service or KWallet. Without one of these, Crewyard refuses to store a key rather than keep it in the clear.

Crewyard 0.3.0 makes agents ask far less, adds goal drafting to the task assistant and gives the Crew graph real actions. It also fixes two problems found while testing 0.2.0 on a real machine.

Fewer approvals

Codex on Windows wraps every command in powershell.exe -Command '…', so in 0.2.0 each read and each git status was an Inbox approval.

• Crewyard now looks inside a runner's shell wrapper and checks each statement of the script separately.
• Inside a step's own worktree, these run without asking: reads and local git (Get-Content, Get-ChildItem, Select-String, rg, git status/diff/log/show, git add, git commit) and the project's own test and build commands.
• Allow always now saves the command that ran, not the wrapper around it.
• These still ask:
  • paths outside the worktree, a .. step, ~ and secret files;
  • pipes, redirection and substitution.
• Pushes, hard resets and package installs stay high risk.

Task assistant and goals

• draft_goal: the assistant can turn a conversation into an Orchestra goal.
  • It fills the composer in Orchestra mode with the goal, a planner and a suggested budget, and adds a goal card to its drafts.
  • Nothing is planned until you press Plan goal, which shows the planning cost first.
• Autopilot (opt-in) can now plan a goal and approve its first plan within the limits you set. Every estimate must be known, each task must fit the per-task limit, the total must fit the day limit and parallel tasks must fit the concurrency limit. Replans and added tasks stay with you (ADR 0022).
• The assistant's instructions now spell out how to change or remove its own drafts, and say who can do what it cannot.

Crew graph

• Task node menu (right-click, Shift+F10 or the Menu key):
  • open details, open changes;
  • retry failed steps;
  • add a follow-up step;
  • pause all, resume all;
  • cancel task;
  • on a proposal: approve, or change the AI of every step.
• Change AI before it starts: for an approved step that has not run yet. Once it runs, Change AI is the existing handoff.
• Run again with another AI: for a finished step. The new step waits for your approval.
• Standby agents are now a small badge on their step (Standby: Gemini CLI +1) instead of full tiles. The badge opens the order, which you can change until the step starts. A real handoff still shows as its own tile with the amber dashed line.

Fixes

• Agent CLIs were shown as not found right after start on a busy Windows machine. The lookup no longer starts where.exe, whose five-second limit could expire under load. It reads the PATH folders itself, in the same order and never in the current folder.
• ChatGPT and Google sign-in were unreachable without an API key: the option sat only in a provider's own dialog, which a provider without a key does not have. The Connect dialog now offers Sign in with ChatGPT instead (or Google) directly.

Release and site

• Plain versions such as 0.3.0 are now published as normal releases, and GitHub marks the newest one Latest. Only tags with a pre-release part, such as v0.4.0-rc.1, are pre-releases.
• The website lists this release on its download and changelog pages within the hour.

Known limits

• The macOS and Linux packages are built and smoke-started on GitHub's runners but have not been tried on a real machine yet.
• No build updates itself yet; download the next release from the releases page.

Crewyard 0.2.0

Install

These builds are not signed with a paid certificate yet, so each system warns once before the first start, and they do not update themselves: get the next version from this page. To check a download, compare the output of Get-FileHash <file> (PowerShell), shasum -a 256 <file> (macOS) or sha256sum <file> (Linux) with its line in SHA256SUMS.txt.

Windows 11, x64: run Crewyard-Setup-<version>-x64.exe. It installs for your user only, without an admin prompt. SmartScreen says "Windows protected your PC" because the publisher is unknown: choose More info, then Run anyway.

macOS 13 (Ventura) or later: -arm64 is for Apple silicon (M1 and later), -x64 for Intel Macs. Open the .dmg and drag Crewyard to Applications, or unzip the -mac.zip. The app is not notarized, so the first start is blocked with a message that Apple could not verify it. Open System Settings, then Privacy & Security, scroll down, choose Open Anyway next to Crewyard and confirm. Or, in Terminal: xattr -dr com.apple.quarantine /Applications/Crewyard.app.

Linux, x64 (glibc 2.35 or newer: Ubuntu 22.04, Debian 12, Fedora 36 and later):

• Debian and Ubuntu: sudo apt install ./crewyard_<version>_amd64.deb. It adds Crewyard to the app menu and crewyard to your PATH, and on Ubuntu 24.04 and later installs the AppArmor profile Chromium's sandbox needs. Prefer it there.
• Any distribution: chmod +x Crewyard-<version>-x86_64.AppImage, then run it. It needs FUSE 2 (sudo apt install libfuse2, or libfuse2t64 on Ubuntu 24.04). Where user namespaces are restricted (Ubuntu 24.04 and later) its launcher starts the app without Chromium's sandbox.

API keys are kept by Windows DPAPI, the macOS Keychain, or the Linux Secret Service or KWallet. Without one of these, Crewyard refuses to store a key rather than keep it in the clear.

Crewyard 0.2.0 is the first public release. The repository is now public under the MIT licence, and this is the first build for macOS and Linux as well as Windows. It is a pre-release: builds are unsigned and do not update themselves.

Orchestra, Crew and Solo

The composer has three ways to run work, and Orchestra is the default when a profile can plan read-only (ADR 0019).

• Orchestra:
  • You write a goal. One read-only planning run, with its estimate and ceiling on the button, returns a plan of up to eight tasks with their dependencies and the files each will change.
  • One click approves it with a prefilled goal budget. Tasks that touch the same files are ordered for you.
  • The scheduler releases tasks in order, at most two at a time per goal. Each finished task merges into the goal's own branch, in Crewyard's own worktree.
  • A last review of all changes opens the pull request gate. You open the pull request.
• Crew: the flow from 0.1: one task, a proposed crew you approve.
• Solo: one agent, started by your click on its shown estimate. It is never remembered as the default.
• Your click is still needed for: a revised plan, a replan, an added task, a research report that code builds on, and spend past any cap.
• Always yours: push, pull requests, merges into your branches and deletes.
• Pausing: Pause goal and Pause all goals hold planning, retries and review loops too.
• New project daily cap. The project's daily budget from the setup wizard is now enforced on every paid run and fallback.

External access over MCP

An agent you run outside the app (Claude Code, Claude Desktop, Cursor, Codex) can connect to Crewyard as an MCP server (ADR 0021).

• What it can do:
  • read projects, tasks, reports, the Inbox, spend and settings;
  • add a folder as a project;
  • tighten project settings;
  • change test and build commands and MCP servers;
  • propose tasks.
• What it never does: approve, start, push, merge, delete, answer the Inbox or read a key.
• Off by default. Settings → External access turns it on, gives each client its own token (kept with DPAPI) and a copyable config, and revokes a client. No TCP port is opened.

Task assistant

• Live state: the conversation shows what the assistant is doing under your message: starting, thinking, or the latest call with the time since the turn began.
• Replies: rendered as text-only Markdown with Copy, Quote and Make drafts from this; code blocks copy.
• Message box: clearer, with a primary Send and a Stop.
• Read tools: the assistant reads much more of the app: tasks, reports, crews, the Inbox, events, spend and limits, profiles, settings, local git, problems and your instruction files (ADR 0018).
• Drafts: it can refine its own open drafts.
• Open task now takes you to the task from anywhere.

Interface

• Bottom panel: it folds to its tab strip from a toggle in its head, like the sidebar and the inspector, and never stops anything when folded.
• Research reports: in the inspector they read as a document, with a severity bar that filters, expandable findings, file:line links into the Files tab, and a Suggested tasks block and tab.
• Colours: red, amber and green now mark the controls and states that mean them: Stop, Turn on and Turn off, risk, verdicts.
• New Guide screen (G W): the three ways to run work, the flow, a getting-started checklist read from your setup, what else Crewyard does, and questions.

Providers

• Gemini CLI's Google sign-in is labelled for Gemini Code Assist Standard or Enterprise accounts. Google stopped serving personal accounts in Gemini CLI on 2026-06-18; use a paid Gemini API key instead.
• New models by API key: Gemini 4 Argon and Gemini 3.8 Flash, with Gemini CLI and OpenCode.
• Claude profiles stay API-key only.

Platforms and project

• Platforms: Windows 11 x64; macOS 13 or later on Apple silicon and Intel (ad-hoc signed, not notarized); Linux x64 as AppImage and .deb.
• Checksums: every release carries SHA256SUMS.txt.
• Licence: MIT, Copyright (c) 2026 Yusuf.
• Docs: SECURITY.md and CONTRIBUTING.md are new.
• Website: lists every release as it is published.

Known limits

• macOS and Linux packages are built and smoke-started on GitHub's runners but have not been tried on a real machine yet. Reports are welcome.
• Orchestra's planning estimate is a fixed heuristic until real planning runs measure it.
• No build updates itself yet; download the next release from the releases page.

Crewyard 0.1.0 (pre-release, unsigned)

First test build of Crewyard for Windows 11. It is a pre-release: unsigned, and built without an update feed.

Before you install

• Unsigned. Windows SmartScreen says the publisher is unknown. Choose More info, then Run anyway. A signed build waits for the certificate decision in PLAN.md §17.
• No automatic updates. This build never looks for updates; install later releases by hand.
• Per-user install. Installs to %LOCALAPPDATA%\Programs\crewyard without admin rights. Your data stays in %APPDATA%\Crewyard when you uninstall. Silent install: Crewyard-Setup-0.1.0-x64.exe /S.
• Requirements. Windows 11, Git, the GitHub CLI for the GitHub features, and the agent CLIs you want: Claude Code 2.1.312 or newer, Codex 0.147.0 or newer, Gemini CLI 0.25.0 or newer, OpenCode 1.19.5 or 2.2.4. The Start screen checks them.

SHA-256 of the installer: ff36079e83e979de3ec2236876ed49ff7da7745bae4773e80a337e168cdc31b6

What it does

• Crews per task. Describe a task and Crewyard proposes who plans, implements, tests and reviews it, with a reason and a cost estimate for each pick. Nothing starts until you approve. In the Crew graph you can edit a proposal (add or remove steps, change roles, profiles and dependencies, order the standby chain), and while it runs you can retry, skip, pause all or add a follow-up step.
• Task assistant and ready-made analyses. A terminal-style chat on a profile you pick drafts tasks and research for you; ready-made analyses cover project review, optimizations, bugs, security, dependencies and licences, and test gaps. A finished analysis shows its report and findings, and suggests follow-up tasks you can add as drafts.
• Four runners, many providers. Claude Code (API keys only), Codex, OpenCode and Gemini CLI over ACP. Providers come from the models.dev catalog: DeepSeek, Kimi, GLM, Qwen, MiniMax, OpenRouter and more. You can pin a catalog version and compare prices against LiteLLM.
• Your own setup. Share the instruction files, rules, skills and commands you wrote for Claude Code, Codex, Gemini CLI, OpenCode, Cursor, Windsurf, Copilot or Cline with your Crewyard profiles. Role files in .crewyard/roles change or add crew roles.
• Limits and spend you can trust. Every number says Reported, Estimated or Unknown. Spend caps apply per key and per task. The sidebar shows the accounts in use first, with details on hover.
• Fallback and handoff. On a limit or empty credit, a task waits, asks, or hands over to a standby in the same worktree, by your rule. Switching between accounts of the same provider is opt-in.
• Inbox and turn control. Questions and approvals with risk levels. You can queue, steer and undo turns, and attach images and files to tasks and messages.
• GitHub under your control. A push queue, pull requests, merge gates, and failing checks you can send to the agent after reading them.
• In the inspector and around it. Files and Preview tabs (a dev-server preview with an element picker), the Activity drawer with events and redacted logs, a taskbar badge for waiting requests, and Open in Crewyard in Explorer.

Known limits

• The task assistant, agent-file sharing, Crew graph editing and the preview screenshot are covered by tests with scripted runners only; they have not run against real CLIs yet.
• Gemini CLI does not report a plan; Crewyard shows the plan as not reported.
• Linux and macOS are prepared but not shipped.

Checked at 151d7e9 on Windows 11: pnpm check (1,654 tests) and 48 end-to-end flows passed.