← All releases

Guildwise HQ 1.0.0

26 October 2026

The first public release of Guildwise HQ, a way to run an organization of AI agents (set up for software delivery by default), under the Apache License 2.0. The repository keeps the development history, with the parts specific to the hosted service removed. **Work and delivery** - Board with cards that move through spec, build, review, test, finalize and QA, with filters for blocked work, rework and items waiting for you. - Project backlog, intake of tickets, and research cards. - Discovery: turn an idea into a PRD with an analyst agent. - Boardroom: a multi-persona discussion run by a chair agent, with minutes and "pull into PRD". - Documents, including PRDs and minutes, with approval and version history. - Phases, follow-ups, a release queue, environments and delivery recipes for deploys. **Agents and team** - Agent roles: business analyst, planner, architect (optional), builder, pre-reviewer, reviewer, tester and DevOps engineer. - Hiring with seniority and per-project allow-lists; pausing, resuming and changing models. - CEO agent in each organization, with a daily report and a desk for routine items. - Org chart, custom roles, departments and workflows. - Ledger of decisions, requirements and standing orders. **Runtimes** - Claude Code, Codex CLI, Gemini CLI, opencode and direct model APIs. **Safety and control** - Desk for questions, approvals and go-aheads; approval gates for risky steps. - Sandboxed sessions: bubblewrap and socat on Linux and WSL2, Seatbelt on macOS, and a separate git worktree per session. - Budgets and cost tracking per project and organization, with CSV export. - Governance: "What it may do" per agent, audit log with CSV export, and an offboarding checklist. - Two-factor login; organizations are isolated from each other. **Knowledge** - Vaults: git repositories of notes, synced to workstations. - Capabilities and packs from git repositories, reviewed before use; playbooks and the `/` menu. - Interests and automations. **Operations** - Workstations: runner for Linux, macOS and WSL2, with a guided installer; native Windows as an explicit, unsandboxed opt-in. - Cloud workstations: a VPS over SSH, or a GitHub Codespace. - Integrations: git hosts (GitHub, GitLab, Azure DevOps, Bitbucket Cloud) and Jira. - Metrics, in-app help articles and a glossary. **Self-hosting** - Docker Compose stack: Caddy for HTTPS, the application, Postgres 18 and a media service for help videos. - Migrations applied on boot. Platform account and organization invitations. - Optional: Web Push, inbound mail, a landing-page assistant and a weekly blog digest. - Every feature is included and organizations are unlimited: a self-hosted install has no plans, trials or limits. - Set `AOM_SECRET_KEY` and keep it with your backups. - Back up the database before every upgrade. See [docs/SELF-HOSTING.md](docs/SELF-HOSTING.md). - Update workstation runners together with the server.

What's new

Accounts for every AI runtime and provider

Integrations now has an AI runtime section: Claude Code, Codex, the Gemini CLI and your API providers, each with a main account and named accounts. Sign in from the browser, set keys, and choose a different account per project in its settings.

One Environments page for where your app runs

Every project now has an Environments page: how to start your app, the test site, the live site, any extra site, and the computers and servers they run on, each listed once. Add as many as you need, remove any of them (the running app is never touched), and let DevOps look first: its questions are real fields you can answer. Apps that run directly on a machine (a service, pm2) work as well as Docker apps.

Deploys is now Releases: what waits for you, what is live on every site, the history and the rules. Everyone can look; the Owner or an Admin releases, and can go back to an earlier version.

Your CEO is now your front door

Talk to your CEO like a colleague on your own machine: in your chat it is a real Claude Code session with your settings, skills and tools (after you run aom-runner ceo full-power on once on that workstation). Pick the model per message or per conversation, and if the workstation is offline the message waits and is sent when it is back. When the CEO needs a permission, you or an Admin get a desk question to authorize or deny.

You can give the CEO full power on one of your own standing orders, pinned to its exact wording. Say "you did X, I'd rather Y" from the chat, the Activity log, the desk or the daily report: it is remembered at once and you can undo or redo the action. The CEO uses the board first, delegates installs and repairs to DevOps (the DevOps chat is gone: ask the CEO), and the daily report shows Things you keep asking me when you repeat a kind of job.

Your CEO now does the reporting; the BA only writes specs

Your CEO now triages what agents raise, watches for trouble, writes your daily report and answers your scheduled tasks. The BA no longer reports to you: it turns requests and tickets into cards. The old setting Your BA reports to you is gone, because the CEO is always there (and Pause on the CEO page is the emergency stop). Items reach your desk unchanged when the CEO cannot run. Reports the BA wrote earlier stay where they are, schedules that asked the BA now ask the CEO, and the report time is on the CEO page. On New intake you can now tell your CEO what you need.

Deploys you approve once, that run without a model

A project can now deploy through a recipe: a small script in the project vault that always does the same steps (deploy, wait until it is healthy, smoke check). Ask your CEO to "set up deploys for this project": it asks about the environments, the targets, the gate and the secrets, and DevOps writes the scripts and tries them. Then you approve that exact version under Project settings, Delivery. Change a byte and it needs a new approval.

The recipe says who releases each environment: you (the new Deploys page lists releases that are ready, with the commit, the environment and the servers, and you press Release, choosing the servers when there are several), the CEO (only where the recipe names it), or automatically once the gate holds. A gate can require an earlier environment to be healthy on the same commit, tests to have passed, a time window and a freeze. A model never runs a deploy command: it can only ask Guildwise, and your workstation runs the script itself, with secrets that stay on that workstation. Projects that deploy with instructions work as before.

A brain for your project

Every project now has a code map: a map of its code and docs, built on your own workstation and kept in the project vault. Agents use it to answer "where is this, and what touches it?" without reading many files, which aims to spend fewer tokens. We measure it honestly: one card in five runs without the map for four weeks, and the Costs page claims a saving only when the numbers support one. Nothing from your code is sent to Guildwise HQ.

Review what your CEO decided

Decisions your CEO takes on its own, such as which working branch to keep, now appear under CEO decisions on the Desk. Acknowledge them (or Acknowledge all), or Correct a wrong one so the CEO learns from it. DevOps questions now go through the normal chain of agents before they reach you, and merge conflicts on a project that does not need your approval for merges are no longer asked.

When you Correct a decision, the CEO and the BA now discuss it: the BA analyses what went wrong and what it caused and proposes fix cards, the CEO replies and records the lesson in its own words, and the BA creates the fix cards. The Desk item shows the discussion as a timeline, with the status "Under review" and then "Lesson recorded · N fix cards". You do not need to do anything.

Turn on notifications

Guildwise now offers to send you push notifications for blocking questions, approvals and runner alerts. Tap "Turn on" when the card appears, or switch them on any time in Settings. On an iPhone, add Guildwise to your Home Screen first.

Pause a project by hand

Owners and Admins can now pause a project from its board. Nothing new starts on it (cards, questions, DevOps, BA work, schedules) until an Owner or Admin resumes it; shifts already running finish, and cards keep their place in the queue. Everyone on the project sees a banner with who paused it and why.

Designs no longer wait for dependencies

A card now counts a dependency as done once its work is merged, even while it waits in Manual QA. The architect can also start designing before the dependencies land: it notes its assumptions and checks them in a short follow-up before the card goes to build.

Edit standing orders

A standing order can now be edited on the Ledger page instead of replaced. Agents use the new wording from their next shift or chat, and History keeps the earlier versions. The DevOps engineer, chats and the BA's background tasks now follow your standing orders too.

Cards are tested again

Cards now always go through Test. With integration tests off (the default) the tester checks the branch against the test plan with what it can run without a running app, as before; with them on it also gets a running copy and your suites. Cards that were going to skip Test now go through it.

Reviewers now try to break your changes

Every review now includes an adversarial pass and a risk analysis. The reviewer rates the change low, medium or high risk and says what else could break and what to test by hand. You see it on the card. A high risk makes the merge wait for your approval on the desk, and the tester and the manual QA steps look at the risky areas first. Reviewers run at high effort, and nothing lowers that automatically.

Run a branch reads your build guide

Run a branch now finds a BUILD.md kept in the folder above your repositories (or the one you name in Run and test), fetches submodules, and checks the environment's own settings before it says something is missing. A failed environment stays on the board with its reason, Try again and Dismiss, and when the DevOps engineer needs an answer from you it shows Waiting for you.

Choose how many shifts run at once

On the Workstations page you can set how many shifts a workstation runs at once, a number or Unlimited, without touching the runner's config file or restarting it. The machine's memory and CPU and the model usage limit still apply.

Work pauses when the Claude limit is reached

When your Claude subscription reaches its session or weekly limit, projects now pause until it resets instead of failing and retrying. The board shows "Paused until 18:00", cards keep their place in the queue, and work resumes by itself.

Approval steps, incoming webhooks, e-mail replies

A workflow can now wait for a person: add A person approves it on the desk in the workflow builder; they approve, or send the card back with a comment. Automations can start from a private web address that other systems call (the secret is shown once), and its fields can fill the card. On a card made from e-mail you can reply to the sender from inside Guildwise, and one noisy sender can no longer use up your inbox allowance.

Archive, spend and quicker card edits

Done cards leave the board a day after they are finished and move to the project's Archive (link on the board): search it, filter it, open a card or reopen it. You can now change a card's priority and completion action at any time, even while an agent works on it. The board shows the real spend this month also when the project has no budget. The BA no longer reads code or sets how a card lands: it hands code questions to the Architect, and only you change the completion action.

Hand the BA a list of tickets

On the Intake page, switch to Batch and paste many tickets, one per line. The BA works through them in the background and creates the cards without a chat per ticket. If it cannot go on with one ticket it asks on your desk, naming the ticket, and the others carry on.

A CEO for your agents

Every organization now has a built-in CEO agent that answers agent questions, keeps cards moving and writes your daily report. There is nothing to appoint and no training period: it decides within your policy from the start, every decision has an Undo, and you can Pause it as an emergency stop. Settings are on Organization, CEO.

Commits are made as you

Every commit and merge made for a project now carries your own name and e-mail, never an agent's or the product's. Set it under Commit as in the project settings, or leave it empty to use the repository's git setting and then your workstation's. Commit messages and pull requests no longer mention cards, agents or Guildwise, and updates to tickets are refused if they do. If no identity is set anywhere, nothing is committed and the desk says where to set it.

An e-mail inbox for every organization

Your organization now has its own e-mail inbox address from the start (Organization > E-mail inbox), with a short note on forwarding your support address to it. On a card made from e-mail, Draft with BA writes a reply for you to read, change and send.

Integration tests run by the tester

Under Project settings > Run and test you can switch on integration tests (it starts off). A card then waits in Test while its workstation starts the card's branch; the tester runs your suites against it and each result shows on the card, with flaky runs marked. The copy is stopped at the verdict, or kept for a while after a failure. Switch it off for a single card in Details. You can also point the tester at a deployed target instead (opt it in under Deploy targets > Testing; test accounts are sealed for the workstation and never stored here).

Connect Claude, git hosts, Jira and MCP servers from the web

The new Integrations page shows every login on a workstation in one place and lets you connect it from the browser: Claude (a subscription sign-in or an API key), GitHub, GitLab, Bitbucket Cloud, Azure DevOps, Jira and plugin MCP servers. Tokens are encrypted in your browser for that workstation, so the server never sees them. Notices such as "Claude login expired" now link straight to the right tile. Needs a runner with web connect; older runners show their state read-only with the terminal command.

Test a card before it is done (manual QA)

A finalized card now goes to QA before it is done: the person who started it gets a desk item and an agent writes test steps (happy path, edge cases, errors, regressions) on the card. They can pass it, skip the test, send it back with a comment, or give it to someone else. A failed test sends the card back to rework and through finalize and QA again, so you no longer reopen closed cards. The update on the Jira or GitHub ticket is held until the card passes or the test is skipped. It is on for every project; turn it off in the project's settings. Cards already in progress follow the setting when they are finalized.

Agent branches stay off your git host

A card's work branch is now pushed to your git host only when the work goes out as a pull request, and then under your branch prefix (for example pieter/DCPL-4276). Agents can no longer push a branch themselves, and the "push the card branch only" completion action is gone: open cards that had it now follow the project's setting.

Run a branch to try it

Press Run this branch on a card, or Run branch... on the board, and the workstation's DevOps engineer starts the project on that branch so you can try it. You see each service, its links and ports, the commit it runs, a live log, and Stop and Extend. It stops by itself after a few hours. Project settings has a new Run and test section for run instructions and services.

Faster, safer branch runs

Starting a branch again now replays the saved plan in seconds, with no DevOps run; if it fails, the DevOps engineer repairs it. The panel has Restart and Update & restart (fast-forward to the branch head), a stopped environment can be restarted for a day, and the services run in the runner's sandbox and keep running when the runner updates itself.

Sign-in problems on a workstation are shown, and work waits

When the Claude login on a workstation expires, its agents now pause instead of failing card after card: the card goes back to the queue without penalty, you get a notice and an e-mail saying to run claude and /login on that machine, and work resumes by itself once it works. Plugin tools (such as Atlassian) that are not signed in are shown as Needs sign-in on the Workstations page and in the project's MCP servers, with the exact steps.

Unstick an agent, and a working branch that keeps up

A shift that hangs can now be freed from the card, the Agents page or the Workstations page: stop the running command, or stop the shift and either re-queue the card or keep it with the same agent. A shift that has gone quiet is also stopped by itself after a while. The local working branch of each repository now follows origin when it is safe, and shows "N behind origin" with a Sync now button when it is not.

Value report and Monday briefing

The Costs page now shows what your agents delivered: cards finished, the hours of work that is worth, what it cost and the return. Every Monday you also get a short e-mail briefing with what waits on you; switch it off in Settings.

Agents read the knowledge vault in chats

Your agents can now read the knowledge vault in chats, in questions they are asked and in the business analyst's tasks, not only while working a card. In a chat about a project they see that project's vault (or the organization vault); in a chat with no project they see the organization vault, but only on the workstation of an Owner or Admin. The vault is read-only for them: anything that belongs in it is saved through the usual commands. If the vault is not cloned on that workstation yet, the agent says so.

Files on cards

Cards can now carry files: screenshots, designs, documents. Add them on the New card form or in the Attachments section of a card (drag files in; images show as previews; up to 20 MB each). Builders, reviewers and testers get the files as read-only copies next to their work, so they can look at the same picture you do.

Faster builds for agents, and.NET works in the sandbox

Agents on the same workstation now share one set of package caches (NuGet, npm, pnpm, pip, Gradle, Maven, Go), so a package is downloaded once instead of once per agent and shift, and packages you already restored on that computer are reused straight away..NET projects can now build and run their tests inside the sandbox; before, dotnet build failed there and agents had to ask the DevOps engineer to build for them. Reviewers can also run commands again on a read-only worktree. Projects that need different.NET versions work side by side: the workstation picks the.NET install that has every SDK, and the Workstations page shows it, with the other tools agents can use, under Agent toolchains. If a tool is missing, ask the DevOps engineer: it installs the tool and registers it for the agents. This needs the next runner update.

The BA can see Jira screenshots and files

When you hand the BA a Jira ticket, it can now download the ticket's attachments through your workstation, look at the screenshots and SVG files, and attach the relevant ones to the cards it drafts. To allow it, run aom-runner jira-connect on the workstation once (the API token stays on that computer); the Workstations page shows whether Jira is connected. Runner 0.1.3 or newer is needed.

Nothing of ours in your repositories

Unsaved edits that hygiene protects now go into a plain git stash on your machine, never to a branch or to the git host. Pull requests go out as <prefix>/<ticket> (for example pieter/DCPL-4276; set the prefix under Project settings, default feature), and merges the runner makes use your own git name.

Jira, Confluence and other plugin tools for your agents

Tools from the Claude Code plugins on your workstation, such as Atlassian (Jira and Confluence), can now be used by your agents. In a project's settings, MCP servers, you (the organization Owner) grant a server to the project, and the agent must also have it in its tool permissions: on the Agents page each agent card has an External tools (MCP) section where you (the Owner) switch a server on for that agent, built-in roles such as the business analyst included. Servers that can write, send or delete carry a warning, so prefer read-only use. The tools an agent called are listed on the card under External tools used.

Reopen a finished card

A Done or Cancelled card now has a Reopen button. Pick the stage (Build by default) and give a reason: the card keeps its branch and open pull request, and finalizing again reuses that pull request.

Safer working branch base

New working branches now start from your repository's real default branch, and shifts starting together no longer clash. If Guildwise cannot tell which branch is the main one, it asks you to set it in Project settings.