Changelog

Every release, on the record.

What shipped, what changed, and what to know before you upgrade. Newest first.

Giljo Agent Message HubNew ProductSeptember 2026

One shared thread, a task backlog, and clean handovers for the AI coding agents you already run.

What's New
  • One thread across harnesses. Claude Code, Codex, OpenCode, and other MCP clients can read and write the same persistent conversation.
  • A task tool for the things agents notice mid-job. Park tech debt, ideas, and follow-ups without derailing the current build.
  • Handovers to the board. A fresh agent session can pick up state, decisions, and next steps without another handoff file in the repo.
  • Free hosted beta. One user per account at amh.giljo.ai. Nothing to install.
v2.1.0 Minor Release September 6, 2026

Work several products at once, run the whole project lifecycle from your coding agent, and get one honest notification model instead of four.

What's New
  • Open more than one product at a time. The dashboard now shows a tab strip for every open product. Switch between them without losing your place, and a background tab shows a live counter when an agent does something there.
  • Every agent tool can name the product it means. Listing projects and tasks, reading the roadmap, and searching memory all accept an optional product, so an agent working across several products no longer worries about you switching tabs mid-task. Creating a project or task without naming a product now asks you to pick instead of guessing, once you have more than one.
  • Run the whole project lifecycle from your coding agent. Stage, launch with a goal in one call, answer a pending approval, un-stage, mark a project superseded, or revive a completed one, all without switching to the dashboard. The dashboard lights up live while you stay in the terminal. Whether a connected agent may launch and approve is controlled under Settings then Security.
  • Setup can bind a repository to a product. Running setup with a product id writes a short marker into your project's CLAUDE.md and AGENTS.md, so future actions in that repository know which product they belong to and stop asking.
  • A banner follows you across the app when an agent needs you. A decision, a mention, or a hand-off now shows wherever you are, says which project or chat is waiting, and clears itself when you deal with it. The Projects tab in the sidebar shows a badge when an agent makes progress on a project you are not looking at.
  • OpenCode joins the supported tools. Setup installs agent templates for it, and the multi-terminal launch command starts it the way it expects.
Changed
  • Products are now Show and Hide, not Activate and Deactivate. Hiding a product never pauses its projects or agent jobs, and you can work on several projects at once inside a single product.
  • One notification model. Each kind of notification has one job, so nothing announces itself twice. Banners ask, desktop pop-ups appear only when the app is hidden and close when the thing they announced is handled, toasts confirm your own clicks, and the bell keeps a quiet count without pulsing. Multiple banners fold into a single strip with a count.
  • The Message Hub is one space instead of a per-product view. Every thread is listed together with a filter that defaults to the product you are viewing. New threads are tagged with a product automatically, and you can retag older ones.
  • The connect wizard tells the truth. A tool card flips green only when that tool really connects from this machine. "Remove tool" really forgets it. Connection commands now work in every terminal, including default Windows PowerShell.
  • Tool names are final. The temporary older names kept for one release are gone, several tools are renamed for what they actually do, and two pairs of tools that asked the same question two ways are now one tool each. Old names still answer for now and point at their replacement.
  • Staging asks how you want a run to work instead of quietly choosing for you. Set it once under Tools then Agents and you will not be asked again.
Fixed
  • 73 fixes across onboarding, the Message Hub, billing surfaces, and agent plumbing. The ones you are most likely to notice are below.
  • Turning on browser pop-ups is one click. A new "Turn on pop-ups" button in settings asks your browser for permission straight away, so desktop notifications work on a fresh machine or profile without any fiddling.
  • Every notice at the top of the screen can be closed, stays closed after a reload, and clears itself the moment the condition it announced is resolved. The "you were mentioned" notice no longer sticks around forever.
  • The Connect page shows which tools are actually connected. Connecting one tool used to turn every tool green. Newer coding tools greet the server differently, and both greetings are now recognised, so the setup wizard notices a connection without a refresh.
  • The onboarding wizard no longer gets stuck. Connect and install checkmarks tick reliably, the review screen shows your agent's full product proposal, the guided tour recovers when a document upload cannot start, and a new account can finish setup even if its trial lapsed first.
  • The Dashboard, Memory Browser, Products, Projects, Roadmap, and Jobs pages update live, including when work is driven from your coding agent. Opening a product in one browser tab no longer switches what another tab shows, and the Tasks list no longer resets your filters when an agent updates a task.
  • Threads you have read stop saying they have something new, a mention near the end of a long message is no longer missed, and you can create threads again when you have more than one product.
  • Sign-in failures explain what actually happened. A wrong password, an inactive account, a rate limit, and a network problem each get their own message, and a brief network hiccup at start-up no longer signs you out.
Security
  • Privileged actions are granted by your account settings alone. A connected AI client can no longer influence its own permission level; only the settings you control decide what it may do.
Heads-up for upgraders
  • Start a new agent session after updating to see the current tool names. Existing sessions keep working until they are restarted.
v2.0.4 Patch August 21, 2026

Agents stay attached to a conversation, agents you run yourself can show what they are doing, and release downloads are named after the product.

What's New
  • Agents you run yourself can show what they are doing. An agent joining a chat from your own terminal used to sit on "Monitoring" however hard it was working. It can now report working, sleeping, or blocked, in the same colours the Jobs board uses. Agents started from the dashboard are unaffected.
Changed
  • Roadmap editing got easier. Name items the way you already see them, such as BE-0001, change one field without resending the rest, and see every problem in a rejected batch at once.
  • Tools state their rules before you hit them. Several tools taught their own rules by rejecting you. They now say up front which fields take a single value, which need updating one at a time, and when overwriting existing content needs an explicit flag.
Fixed
  • Agents stay attached to a conversation. Waiting for a reply used to end in a timeout on most AI clients, and the agent dropped off the thread. The hold is now shorter than the client limit, and an empty wait says in plain words to keep waiting.
  • Messages to finished agents no longer pile up as phantom unread badges, and you are told when a message could not be delivered.
  • Agent badges agree with themselves. The same agent shows the same initials and colour everywhere, and a name with a bracket no longer renders as "R(". New agent names can no longer contain punctuation.
v2.0.3 Patch August 19, 2026

Your AI assistant can now read a big board without missing a row, search it by a word, and answer an approval without leaving your client.

What's New
  • Approval requests can be answered right in your AI client. When an agent pauses to ask you to approve something, clients that support the newest connection standard show the choice inline and the agent carries on the moment you pick. Other clients are unchanged: the request still waits on your dashboard.
  • Search for projects and tasks by a word. Ask for "the OAuth one" and your assistant finds it by a word in the title, the description, the short code, or the task number, instead of reading the whole list to look for it. Memory search now covers commit messages too, so "when did we fix the redirect bug" has an answer.
  • Search inside a conversation. Open any chat in the Message Hub and the new search box narrows it to what you are looking for, matching both what was said and who said it.
  • Agents can wait for their turn instead of polling for it. A new tool lets an agent hold one call open until a message or hand-off actually arrives for it. Work reaches the right agent in under a second, and waiting costs nothing while it waits. Polling still works exactly as before.
  • See at a glance who on a conversation is still working. Agents can now check, for every participant in a thread, when they were last active and whether they are active, quiet, or gone. An orchestrator can tell a busy agent apart from a stalled one without digging.
  • Your agents can promote a task to a project on their own. An agent converts it in one step and gets exactly what the dashboard wizard produces: subtasks and the roadmap card move across, and the new project arrives inactive, so you still choose when to launch it.
  • Your orchestrator's personality can differ per product. Customise the orchestrator prompt for one product and it applies only there. Products you have not customised keep using your all-products prompt, and the editor now tells you which of the three each product is actually running on.
  • Hand-offs to you survive a reload and follow you between devices. When an agent hands work to you, the notification lives on the server rather than in whichever tab was open, so you find it after a refresh, on a second computer, or the next morning, and it names the agent that is waiting.
  • Codex worker lanes can run headless without stalling. An orchestrator running Codex workers without a terminal for each one can drive them through a new built-in helper, so a lane no longer sits waiting on a confirmation prompt nobody can answer.
Changed
  • One clear "Action needed" style for everything that needs you. A hand-off, a mention, and a request for your approval now look and behave the same in the banner, the bell, and desktop notifications, each with a small label saying which it is. Clicking any of them takes you to the exact message, not just the thread, and mentions and approvals stay in your bell until you deal with them.
  • Your agents follow the product you are working in. Enable or disable an agent and that choice is now remembered per product: it decides what installs, what your orchestrator can start, and what it sees on its roster. Orchestrators can start any of your enabled agents, up to 16, instead of the old cap of 8. A product you have never customised keeps all of your agents.
  • Your assistant sees how big your board is before it reads it. Listings come back with the totals for the whole board and are properly paged, so a shortened answer says so plainly instead of looking complete. A new compact listing view returns just what is needed to find and sort work, typically 35 to 40 percent smaller.
  • Clearing "waiting on you" is a button you can actually find. The old text link at the bottom of the thread is now a yellow hand button that gently pulses while a thread is waiting on you, with a second copy that stays on screen beside the thread search. It respects your system's reduced-motion setting.
  • Agent crews start with the right instructions once, not twice. In Subagent mode each agent already loads its role from the file installed on your machine, so the server no longer sends the same role a second time. That leaves noticeably more room in every agent's context for the actual work.
  • Every agent briefing says where its orchestrator instructions came from: a custom prompt saved for this product, one saved for your whole account, or the built-in default, with the date you saved it. An agent behaving unexpectedly now takes one glance to explain.
  • Check-in cadence is one account-level setting. The per-project slider is retired. Agents on a harness with live wake support respond to new work instantly instead of sleeping on a timer, and the dashboard tells you whether an agent is waiting for a wake signal, sleeping on a countdown, or has gone quiet. Cadence values you had set on individual projects are still honoured.
  • Giljo HQ now runs on version 2.0 of the Model Context Protocol SDK, and supports the newest protocol version alongside the older ones. Your existing connections keep working exactly as before, with nothing to reconnect or reconfigure.
  • Message Hub posts always carry an explicit author. Agents identify themselves on every post, and posting in your voice takes a deliberate flag, so a forgotten field can no longer make an agent's message look like you wrote it.
  • The in-app Privacy Policy and Terms describe the service as it actually runs, naming every subprocessor behind the hosted edition, where your data and backups are stored, and the full retention picture. Self-hosted installs see provider-neutral wording.
Fixed
  • Paging a long list no longer skips rows. When several items were saved at the same moment, walking a list page by page could quietly leave one out, with no error and no warning. Pages now advance on a stable order, so every item is returned exactly once, and a stale page marker says so instead of silently handing back the first page forever.
  • Finished work no longer vanishes from what your assistant reads. The safety cap on the project list kept whichever projects were newest, so a project started months ago and finished last week fell outside the window entirely. Finished work is now kept by when it was finished, unfinished work is never dropped, and a shortened list explains how to narrow the question.
  • A mistyped filter tells you, instead of pretending nothing matched. Filtering by a status like "in progress" used to come back silently empty even when matching work existed. You now get a clear message naming the valid values, and an empty list explains when completed work is simply hidden, and the exact option that reveals it.
  • Installing your agents can no longer overwrite files you wrote yourself. Every exported agent now records that it came from Giljo HQ and which product it belongs to, so an install refreshes its own files, leaves anything you hand-wrote alone, and asks before replacing anything it is unsure about. Exported agents are named after their product, so the same agent used by two products no longer overwrites itself.
  • Editing a task now saves. Reopening a task and changing its title or description used to fail with a generic "Failed to save task" and lose the edit.
  • Two tasks in the same product can no longer end up with the same number. The database safeguard behind task numbering had quietly stopped working. It is back, and an existing duplicate pair is repaired automatically on the next start, with nothing deleted.
  • Connected AI clients no longer lose their connection when their sign-in token renews in the background, so a long-running session keeps working instead of stopping with an error until you sign in again.
  • Connecting Claude to your server works again. Approving the connection could fail with "Invalid authorization request parameters" and never finish, for any app that does not name a specific target when it asks for access.
  • The Privacy Policy and Terms pages open without signing in. A direct link used to bounce you to the welcome screen instead of showing the document.
  • Connecting from OpenCode works with the copy-paste command. The connect screen was handing OpenCode a command written in another tool's syntax, so the first attempt always failed. Every connect command now names the tool it is for.
  • Pages no longer get stuck failing to load after an update. A file requested during the short window while a new version was going live could be remembered as missing for up to a year, leaving a page blank until someone cleared the cache by hand.
  • Waiting for your turn works on hosted installs, not just self-hosted ones. On a deployment running several server processes, a waiting agent could miss a message that landed on a different process. Self-hosted installs were never affected.
  • Marking a project completed from an agent finishes it properly. It now runs the same full close-down the Archive button does, instead of a halfway version that looked finished on the dashboard while leaving its helper agents hanging around.
  • Tasks and projects created by agents land on the product the agent intended, even while you switch products in the dashboard, and editing one of your own projects no longer depends on which product you have selected.
  • Long messages, missions, and descriptions appear instantly in every open window, instead of arriving only in the window that sent them and leaving other sessions to refresh.
Security
  • Chats can no longer be filed against another workspace's product or project. When a chat was created with a product or project named explicitly, that name was stored without checking it belonged to you. Naming one from another workspace is now refused outright. Chats you create normally are unaffected.
  • Bundled libraries updated to close published advisories. nanoid and dompurify were raised to their patched releases, and the build now uses newer undici, postcss, and brace-expansion releases that resolve published advisories, including a denial-of-service issue. No features or behavior change.
Heads-up for upgraders
  • New agents arrive ready to configure, not switched on. Adding an agent, or pressing "Add Default Agents", used to put it straight to work in the product you had open while it still carried its stock instructions. New agents now appear in your agent list to tailor first, and go live in a product only when you switch them on there. Agents you had already enabled are untouched.
  • Every project now lives under a product. Creating one without naming a product is refused up front with a message saying what to pass. Older installs that still held product-less projects get them filed under a product on upgrade. Nothing is discarded.
v2.0.2 Minor Release August 5, 2026

GiljoAI MCP is now Giljo HQ, and the Message Hub is rebuilt around one question: who is waiting on you?

What's New
  • A raised hand marks every thread that's waiting on you. Yellow now means exactly one thing: an agent handed you the turn. The hand, the gold frame, and the strip at the top of General always agree. Unread threads no longer light up the whole board.
  • You're told it's your turn no matter where you are in the app. The notification reaches you on any page, not just when the Message Hub is already open, and clicking it takes you straight to that conversation.
  • "Mark handled" clears a thread that's waiting on you without posting a message. The thread stays open; the hand and the frame simply go out.
  • Agents can ask you directly for a decision. An agent that needs your call addresses you by name, so the request shows up as your turn instead of disappearing into the conversation.
  • Chats show which tool each agent is running in: Claude Code, Codex, Gemini and the rest, detected automatically on connect, alongside when each agent was last active and what it's doing right now.
  • Rename a chat, and close it yourself. Chats could only be named at creation and only an agent could mark one resolved, so the list filled up with conversations stuck on "Open". Project chats keep the project's name and stay with its history.
  • The chat list tells you what you need at a glance: who has joined, from which tool, the last thing said, and whether anything is new since you looked. It now loads in a single request instead of one per chat, so long lists appear noticeably faster.
  • A focused, 11-tool view of the server. Adding ?profile=listing to your MCP server URL keeps the everyday essentials and hides the multi-agent orchestration tools, so a new connection is quicker to learn. A profile can only ever show fewer tools, never more.
  • Park a project to set it aside without cancelling it. Parked projects drop off the roadmap and stay out of your way; nothing is lost, and reactivating brings them back.
Changed
  • The Message Hub is a calmer place to read. Threads are cards carrying the three things worth knowing: the name, who has checked in and from which tool, and the last thing said. Rename in place, copy the id to hand to an agent, or delete. No menus. Cards no longer grow tall when six agents join.
  • Conversations read like conversations. Message text renders formatting, a run of posts from the same agent groups under one name, and a very long post folds to its first line until you ask for the rest.
  • Less noise on screen. The message counters, the waiting/read/sent filter row, and the "broadcast" tag that appeared on nearly every message are gone. A tag now appears only when a message really was sent to one person.
  • Your default agents now speak in terms of your project. The tester, implementer, and documenter templates no longer assume a specific test framework, coverage target, or multi-tenant setup, and the agent protocol no longer hardcodes our own toolchain, Python-only code navigation, or a Claude-Code-only task tracker. Instructions are derived from your product and the tool you're actually using. Applied automatically on your next update, including for existing accounts.
  • MCP tools now show clear names and safety hints in connector directories. Every tool exposed over MCP carries a short, human-readable name and an accurate "read-only" or "changes data" hint, so clients like Claude and ChatGPT can warn you before a tool makes a change.
  • Connecting a tool happens in one place. The API Keys card is now purely for viewing and revoking keys, the duplicate "Configurator" window is gone, and the Generic MCP client card gives you a copyable connector URL: paste one link to connect a web app or IDE, no API key needed.
  • Each of the four ways to create your first product tells you what it expects before you pick it: whether it needs a connected agent, roughly how long it runs, and who is doing the talking. Both agent-driven paths now say something useful when nothing comes back instead of waiting indefinitely.
  • The certificate trust steps are now where you need them. If your AI coding tool refuses to connect over HTTPS, the walkthrough is reachable straight from Tools → Connect and from the connect card itself, not only from Startup.
  • Chain project cards show a Planning badge while a project is being staged, so a project actively underway no longer looks identical to one still waiting in the queue.
Fixed
  • Hand-offs no longer go missing. Addressing an agent by its display name instead of its id used to send the message and the turn to nobody, the commonest way a hand-off vanished. Both are now checked before sending, the refusal names the exact id to use, and passing the turn to someone who cannot receive it is declined on the spot instead of reporting success.
  • A message to an agent that already finished is no longer lost. It's automatically forwarded to the agent still driving the project and you're told right away who picked it up; if that hand-off fails to save, you're told that too.
  • Agents in the Message Hub are no longer mistaken for you. An agent identifying itself with an id could be shown in your colour, on your side of the conversation. Identity now comes from the server instead of being guessed from the shape of an id. Everyone who speaks in a chat appears in its participant list, an agent's name sticks once it introduces itself, and its detected tool is no longer forgotten.
  • Multi-project chains no longer lose (or duplicate) their coordination chat. Each chain's chat is linked directly to the run instead of being found by its title, so agents stop going silent when a title changes, and a restarted chain is pointed back at the chat that already exists rather than opening a rival one. Existing chains are relinked automatically.
  • An agent that goes quiet after finishing its work can be accepted, not written off. Record what it delivered and close it out normally; it shows as closed instead of being filed with agents that failed. Opening a stalled agent's task list no longer counts as the agent waking up.
  • Deleted things stop haunting the dashboard. A deleted product's 360 memories, git commits, and name no longer appear beside your projects; deleted projects no longer inflate project totals or tag counts; and a deleted agent can no longer be spawned, exported, or offered when starting new work.
  • Two housekeeping jobs that had quietly stopped now run again. Resolved notifications are cleared on schedule, and items you deleted more than 30 days ago are permanently removed as intended. Both had been failing silently since an internal safety check was tightened; nothing was ever lost or exposed, the cleanup simply never happened. Self-hosted installs are the ones affected.
  • Setup now works when you connect from ChatGPT. Hosted chat connectors used to get file-install instructions for folders they cannot write to. Those sessions now receive their agent templates directly in the response, in plain Markdown. Desktop and terminal tools keep the normal file install.
  • Linked projects now run in the execution mode you picked for the chain. Projects set up individually and then linked used to keep their own mode, in some cases handing one project two conflicting sets of instructions. The chain's mode now decides for every project in it, and locks once a project has launched.
  • Your context depth settings no longer reset themselves, and the tech-stack detail setting now genuinely changes how much your agents see instead of always sending everything. An unrecognised depth option returns a clear error instead of quietly falling back to the default.
  • A busy server no longer signs you out or throws you into the first-time setup screen. A temporary rate limit is treated as "try again in a moment" rather than an expired session or a brand-new install.
  • The Completed date on your projects is the real completion date, not "last modified", so archiving or editing a finished project no longer changes it. Filtering and sorting by completion date work again, and existing projects were backfilled from their closeout records.
  • Starting work from the command line shows up in the dashboard straight away. Staging from a CLI or headless session used to leave the orchestrator card missing until you navigated away and came back.
Security
  • Sign-in tokens are now encrypted while briefly held in the server-side cache. No readable token ever sits in the cache, closing the path where cache storage or monitoring could have exposed live credentials on hosted deployments.
  • Connecting an app is hardened end to end. A one-time authorization code can only ever be exchanged once (present it twice and any tokens already issued from it are revoked immediately), the proof tied to that sign-in is always confirmed, and revoking a connection invalidates its access tokens at once instead of letting a revoked session run out the clock.
  • Rate limits can no longer be sidestepped. Login, registration, and connection throttling applies to every request and can't be disabled by a crafted header, and the endpoints that exchange, refresh, and revoke app-connection tokens are now rate limited too. Self-hosted localhost use is unaffected.
  • The last known gaps in workspace data isolation are closed. Notifications, roadmaps, roadmap items, chain runs, and Hub threads and their participant lists are now covered by the same automatic safeguard as your other data, and a bulk write the guard cannot scope to your account is blocked rather than merely logged.
  • The hosted service refuses to start without its configured public address, so sign-in and connection URLs always come from trusted configuration and never from request headers a malicious party could influence. Self-hosted installs keep their flexible reverse-proxy behavior.
  • Dependency and packaging hygiene: the library protecting sign-in tokens and a bundled YAML parser were updated to versions clearing published advisories, the pre-release scan for internal network details now covers the Windows and Linux installers, and log downloads serve only files that genuinely live in your log folder.
Heads-up for upgraders
  • GiljoAI MCP is now Giljo HQ. Reconnect your AI tools with the new giljo_hq server name from the Connect page. Tools registered under the old giljo_mcp name keep working until you re-add them.
  • The public repository is now github.com/giljoai/giljo-hq. Existing links and clones redirect, so nothing breaks, but update your bookmarks.
v2.0.1.1 Patch July 17, 2026

Your agent can build your first product for you, guided by a new animated welcome tour.

What's New
  • Your AI agent can now create your first product for you. Two new tools let a connected coding agent create a product and write its vision document directly, so onboarding is one prompt pasted into your CLI while the product card fills itself, instead of typing everything into a form. Agent-written vision documents appear in the dashboard exactly like uploaded ones.
  • A new animated welcome tour. After setup, five short animated steps walk you through your tools, your product, your agent crew, missions, and memory, instead of a wall of text. It ends by asking how you want to start: import an existing codebase, shape a new idea, upload a vision document, or fill the form yourself. It then carries you to your first activated product. Your progress survives a reload.
  • Review a completed project straight from its notification. Closeout notifications open the project's Implementation tab, where a new "Review project" button reopens the closeout summary: outcomes, decisions, and commits.
  • One-click "Add Default Agents." Bring back the built-in agent templates at any time. The import only ever adds: templates you edited are kept as they are, and defaults you already have are skipped.
  • A gentle context-review reminder. When your active product hasn't had its context reviewed and you've completed work since, a banner suggests tuning it so your agents keep building with an up-to-date picture. Dismissable, and off entirely if you've turned it off in notification settings.
Changed
  • You can now keep up to 16 agent templates active at once (was 8), up to 15 of your own alongside the reserved Orchestrator, reflected in the "Your Team" badge and the Agent Template Manager.
  • Agents can hand over the turn in the same call that asks the question. A message that requires action now moves the thread's turn with the post, so the recipient's "is it my turn" check can no longer miss a question that forgot the separate hand-off step.
  • A clearer way to connect your coding tools. The setup wizard walks you through your tools one at a time with a live status card that turns green the moment a tool connects; nothing to click. Six tools including OpenCode and any generic MCP client, each with its exact one-command setup, and your connected tools now live in a single directory with live status.
  • The in-app User Guide caught up with the app. Every chapter was reconciled against the shipped product: new chapters for the Message Hub and the Roadmap, an accurate walkthrough of the redesigned setup wizard, chain projects, the welcome tour, backups and account security, and a "Limits at a glance" table with the real numbers. Retired content is gone.
  • Release notes now ship with every change, enforced automatically. Each improvement records its own release-note line as it is built, and an automatic check keeps the set complete, so nothing is left behind at release time.
Fixed
  • Long agent names no longer break the Message Hub or chain runs. Broadcasting, handing off, or directing a message on a thread with a long-named participant works reliably, and chain runs accept a long conductor name.
  • Agent setup downloads always reflect your current agents. A setup link now refuses to hand back a stale bundle, so you can never install a snapshot missing templates you just added.
  • Your own agents are never crowded out by the built-in defaults in an agent export.
  • The agent template page's filters work. The status filter offers only the states templates actually have, the category dropdown filters by agent role, and the Export Status column is sortable so out-of-date templates are easy to find.
  • Clear message when a product or project name is too long: a friendly validation message with a live character counter instead of a server error. Uploading a vision document whose name already exists now says so plainly, and retrying is safe.
  • A friendly "page not found" screen replaces the unfinished placeholder, with buttons to go home or go back.
Security
  • Force logout now ends connected app sessions too. Previously a connected app could keep refreshing its own access after a force logout.
  • Closed a timing gap where a connected app could survive being logged out. If a session was invalidated at the exact moment an app was refreshing, the app could slip through. Refreshes and invalidations are now serialized so the invalidation always wins.
  • Login timing no longer reveals whether an account exists. Sign-in takes the same amount of time either way.
  • Rate limits can't be bypassed behind the proxy. Login and password-reset throttling now always keys on the real client address, closing a gap where a forged forwarding header could dodge it.
v2.0.1 Patch July 15, 2026

Bulletproof backups: your data is captured automatically, forever.

What's New
  • Backups that maintain themselves. Account backups, restore checkpoints, and the GDPR data export now derive their coverage directly from the database schema. Every feature we ship, present and future, is captured in your backups by construction, and an automated guard fails our build if anything is ever left out. Backup files are version-stamped, and older backup files keep restoring cleanly.
  • The "Messages Waiting" badge is back on the implementation page. Better still, it now updates live as your agents receive and read messages, so you can see at a glance who has mail.
Changed
  • Message Hub tabs renamed for clarity: "Project Comms" is now Project threads, and "Town Square" is now General threads.
  • Dependency refresh across the stack: FastAPI 0.139, MCP SDK 1.28, Vite 8.1.4, Vuetify 4.1.5, and current versions of dompurify, marked, eslint, prettier, the Sentry SDKs, and js-yaml.
Fixed
  • Account restore could fail (or silently drop Message Hub data) for recent accounts. Backups now capture hub threads, participants, and read state; restores work again for current data, and older backup files restore with every message kept.
  • Signed-in sessions no longer hang after a failed token refresh: queued requests settle cleanly and you land on the login page instead of a frozen app.
  • Project closeout could deadlock when an orchestrator was never staged; force-close now frees it, and fully-completed states route straight to closeout.
  • Several agent-tool errors fixed: task lists filtered by due date, task due-date updates, and context reads with unusual parameters no longer return internal errors.
  • Very long passwords are rejected with a clear message instead of a server error.
  • The Admin/Owner badge no longer shows on hosted accounts (it's meaningless on a single-user workspace).
Security
  • Sensitive account changes now require a live browser session. Email, password, and recovery-PIN changes, along with creating new API keys, can no longer be driven by an API key alone.
  • Cross-tenant hardening pass: WebSocket connections are not discoverable across workspaces, task references are verified to belong to your workspace before use, and hosted accounts pin the request host against spoofing.
  • Log and telemetry hygiene: authentication tokens no longer appear in access logs, error reporting scrubs sensitive request details, and the final log-injection findings from our public code audit are closed.
  • Password checks fail closed if the hashing backend ever misbehaves.
v2.0.0.2 Patch July 14, 2026

Post-launch hardening: a clean bill of health from the public security audit.

Security
  • Log-forging protection completed across the backend. Every place user- or agent-supplied text reaches a log line now strips control characters, closing the log-injection findings from the public security audit.
  • The public health endpoint no longer reveals internal error details during an outage; full diagnostics moved to server logs and the authenticated status endpoint.
  • Startup lock file is readable by its owner only, protecting startup coordination on shared machines.
Fixed
  • Vision-document import tuned for accuracy: cleaner, more faithful extraction when importing a vision document.
  • The vision-analysis wizard no longer gets stuck on "Analyzing" if a realtime update is missed.
  • Vision analysis fills in the Codebase Folder automatically when it can determine the project path.
Removed
  • Operator-only dev tools no longer ship in Community Edition.
v2.0.0.1 Patch July 13, 2026

Installs behave identically on every locale.

Fixed
  • The installer no longer fails on non-UTF-8 Windows systems. All installer-generated files are written explicitly as UTF-8 and the configuration template is pure ASCII, so first-run setup now works the same on every regional setting.
v2.0.0 Major Release July 13, 2026

The relaunch: your AI team gets a conductor, a message hub, and a memory.

What's New
  • Chain projects: link projects and run them back to back. Select several projects, link them into a chain with a shared mission, and a dedicated conductor stages each one, launches it, watches progress, and advances to the next as each completes. Review checkpoints and a firm human-in-the-loop halt before implementation keep you in control while the Jobs view follows the whole run live.
  • Work from chat-based AI tools and web coding agents, not just terminal CLIs. GiljoAI now identifies the connected AI tool and adapts its instructions to what that tool can actually do. Web coding agents can hand work across pull requests, and setup includes presets for the latest AI coding tools.
  • Message Hub: a built-in messaging center for you and your agents. Conversation threads bind to projects and jobs, agents post under their own identity, and the Jobs view opens each job's thread directly. Coordination happens in your workspace, visibly, not in invisible side channels.
  • Sign in with Google or GitHub. Sign up and log in without a password, connect a provider to an existing account, or add a password to a provider-linked account any time.
  • Fully self-service Hosted subscriptions. Subscribe monthly or yearly through a secure embedded checkout, switch plans, cancel or resume, and open your billing portal for invoices, with clear lifecycle emails for payment events.
  • Roadmap pane. A dedicated view that keeps upcoming work ordered: agents maintain it as they plan, you edit and reorder, and the Projects list can sort in roadmap order.
  • Search your 360 memory. A full-text Memory browser across cross-session learnings, plus a search tool agents use to look up prior learnings mid-run.
  • Trash and recover. Deleted threads, tasks, vision documents, and agent templates land in a recoverable trash with scheduled purge. Hiding a project is now called Archived, and search finds archived items.
  • Account backups and restore (hosted). Nightly encrypted snapshots plus Danger Zone controls to download your latest backup, create a restore point on demand, or request a restore; restores preserve your API keys so integrations keep working.
  • Complete, automated account deletion (hosted). Deleting your account auto-cancels an active subscription, offers immediate deletion, and purges everything including backups, with hash-verified deletion receipts retained for compliance.
  • A public status page (hosted). Service health and incident history, linked from the app.
  • New agent tools and supervision controls. Agents gained tools to stage, implement, launch, diagnose, and chain projects; server-enforced tool profiles bound what each connection may call; and you can request automatic agent check-ins on an interval you tune.
  • Manage your own credentials. Change your password and recovery PIN from your profile; hosted accounts can change their sign-in email with confirm-before-switch verification.
Changed
  • Execution modes radically simplified. Six overlapping modes collapsed into two, with the right agent harness detected automatically from the connected tool. Headless running is an explicit toggle, implementation defaults to a human-in-the-loop gate, and the dashboard auto-follows a headless run live.
  • Product creation is AI-first. Attach vision documents and run vision analysis to fill in your product profile: analysis is the default path, documents are stored safely in the database, and completing analysis unlocks staging.
  • Simpler self-hosted install: Community Edition runs over plain HTTP. No certificate step, no HTTP-vs-HTTPS questions; HTTPS is an optional post-install upgrade in Settings → Network with your own certificate.
  • Faster everywhere. Project lists paginate and search in the database, the frontend dedupes requests and pauses polling in hidden tabs, hot backend paths moved off the event loop, repeated-query hotspots were batched and indexed, and connected AI tools receive much smaller instruction payloads. Long sessions and large workspaces feel it most.
  • Refreshed look and feel. Material Design 3 across the app, unified banners, token-driven agent colors, a redesigned Tools → Connect and onboarding, and a compact Projects view.
  • Connecting AI tools is clearer. Guided per-tool connect flows with copy-ready snippets, an OAuth-first connect step on hosted accounts, a streamlined API-key flow for Community Edition, CLI OAuth sign-in, and stay-signed-in via rotating refresh tokens.
  • Notifications are now database-backed. They persist across sessions and devices, deep-link to their project, and consolidate the previously scattered banners, including API-key expiry reminders.
  • The in-app user guide covers the whole product, edition-aware for Community and hosted users.
  • Network settings show the real server address, with a one-step bring-your-own-certificate flow.
  • Projects can be marked superseded with a link to their successor; history stays navigable without deleting anything.
Fixed
  • Live updates are far more resilient: reconnect storms and cold-start outages fixed, heartbeat supervision added, reconnect re-arms when your machine wakes, and the agent transport is hardened against repeated-request storms and oversized payloads.
  • Dashboard navigation no longer stalls, and after a server update the app recovers from stale cached assets on its own.
  • Serial numbers stopped jumping to five digits.
  • Stuck projects can always be recovered: deactivate resets a never-launched orchestrator, unstage releases stale locks and missions, promoting a task no longer deactivates your active project, and refreshing agent templates preserves your edits.
  • Installer and first-run hardening on every platform: safe re-runs with a new repair mode, atomic extraction, pinned dependencies, Windows PATH and prompt fixes, Linux apt-lock patience and WSL guards, and fresh-install seeding fixes.
  • The landing page's "Forgot your password?" link works again (hosted).
Security
  • Credential changes end live sessions. Changing or resetting a password evicts all sessions and refresh tokens, deactivating an account closes its connections immediately, and access tokens rotate on refresh.
  • Stronger sign-in protection: per-account login lockout, failed-attempt throttling on API-key and WebSocket paths, higher password-hashing cost, and uniform auth errors that don't reveal whether an account exists.
  • Workspace isolation enforced in depth: bulk operations pass through the isolation guard automatically, isolation wiring is verified at boot and fails loudly if missing, agent-tool authorization fails closed, and a permanent isolation test gate runs on every build.
  • Boundary validation and rate limiting hardened: agent-supplied input is validated at the boundary with sanitized errors, structured payloads are validated on every write, and the rate limiter is atomic under concurrency.
  • Dependency refresh across the stack clearing all known high-severity advisories.
Removed
  • Built-in statistical summarizer removed. Your connected AI agent produces summaries now: better results, lighter install.
  • The separate demo edition retired: folded into the Hosted tier.
  • Built-in certificate generation removed from the installer: bring your own certificate and add it in Settings → Network.
  • Dead weight deleted: dozens of unused routes, retired setup remnants, and obsolete presets.
Heads-up for upgraders
  • PostgreSQL 16 is now the minimum (18 recommended). Community Edition installs on PostgreSQL 14/15 should upgrade the database server before updating.
  • Community Edition now serves plain HTTP by default. If you relied on installer-generated certificates, switch to the bring-your-own-certificate flow in Settings → Network.
  • v2.0.0 is a fresh public release line. Update via the standard installer; your data migrates automatically on first boot.
v1.3.0 Minor Release May 20, 2026

Live working-time on every job, and a smoother staging-to-implementation handoff.

What's New
  • Working-time on every agent job: see live how long each agent has been actively working, on the agent card.
  • Native spellcheck across the dashboard for descriptions, missions, and messages.
  • Roomier 360-memory summaries: per-entry cap raised from 500 to 1,500 characters.
  • Customize-prompt panel reordered: the customize box sits above the copy button, and your custom text is included in what's copied.
Changed
  • Cleaner staging-to-implementation handoff. Staging finishes into a clear Waiting state until you click Implement: no flicker, no stale labels; implementation kicks off the instant you decide.
  • Smoother orchestrator behavior across phases: one orchestrator per project, transitioning cleanly through Working → Waiting → Complete.
  • Sharper AI-agent integration. Clearer guidance, lighter payloads, cache-aware reads: faster, less chatty agent interactions.
  • Projects sorted newest-first by default.
Fixed
  • Closeout button appears where it should on implementation-phase projects; the closeout spinner stays visible during normal retries.
  • Live message-audit statuses update in real time.
  • Hidden tasks no longer leak into the Tasks view; project-open clicks only fire on the Serial badge; date filters stopped crashing on time-zone edge cases.
  • Setup polish: clearer API-key copy hint; setup templates escape special characters safely; quieter server logs.
Security
  • Frontend and Python dependencies refreshed (dompurify, marked, axios, Sentry SDKs, and more).
Removed
  • Legacy deprecated agent tags fully retired: clear validation errors instead of silent warnings.
  • Unused columns and obsolete tests pruned.