Changelog

What's new in Bunderhost

Every change people can see, newest first, with a picture of the part of the product it touched.

NewMCP

Agents read your product analytics and session recordings

Every production site already counted its visits and recorded a sample of them, but an agent could not reach any of it. Now it can: get_analytics answers about twenty reports, from an overview against last week to pages, query strings, campaigns, custom events, journeys and retention, and get_analytics_funnel shows which step of a flow people abandon. These are aggregates and come with read access.

With visitors:read, which you approve like other customer data, an agent lists the sessions behind a drop-off and reads their recordings. A recording comes back as a timeline of pages, clicks and inputs, never as video and never with page text or what anyone typed, with signals where a visitor seemed stuck: repeated clicks, a click that did nothing, a switch flipped back and forth, a form left half-filled.

Credentials in URLs, such as password-reset tokens and sign-in codes, are now redacted in the browser before a visit is sent, and in everything an agent reads. Production deployments and approved broadcasts appear as notes on the charts, so "did the fix help?" has an answer there. See Product analytics in the docs.

FixedDeployments

Projects with a long release history list their deployments again

A project that had shipped a couple of hundred releases showed an empty Deployments tab, and an agent asking for its deployments, or for the projects of its organization, got an error instead. Every list read each release together with its whole blueprint, and for a long history that came to more than the database answers at once.

Lists now leave the blueprint out and read it only for the one release that needs it, and the command line looks a commit up without reading the rest. There is nothing to do: the full history is back.

FixedPreviews

Previews of closed pull requests are removed again

A preview whose pull request was merged or closed while Bunderhost missed the webhook should be removed by the nightly sweep. Since October it was not: deleting a preview's files failed, because a preview's bucket is a copy of production's and still lists production's files, and that one failure stopped the whole sweep. Those previews stayed listed with their database and files.

The sweep now deletes a preview's bucket the way the storage provider intends, which leaves production's files untouched, and one preview that cannot be removed no longer holds up the rest. The previews left behind are removed on the next nightly run; there is nothing to do.

NewServers

Move a project from Cloudflare to your own server

A project that ran on Cloudflare Workers used to stay there: the deployment target in Settings was read-only. Now you choose one of your servers there, or run bunx bunderhost target fsn-1, and the next production deploy starts the project on that server and switches production. The database stays where it is. A Workerstack app runs under celld; commit a Bunderstack blueprint in the same change and it runs as a container instead.

A custom domain in Cloudflare is held by the Worker, and its DNS record cannot be changed while it is. So once the server serves, Bunderhost detaches the domain and writes a DNS-only record for the server in its place. Expect a few minutes without the domain while the server issues its certificate. When the record cannot be written, or another one is already there, the deployment log names the record to add.

Domains served through Cloudflare for SaaS cannot move yet: remove them, move production, and add them again. Static sites stay on Cloudflare.

ImprovedMessaging

Worker apps choose their email provider on the Messaging tab

The Email delivery switch on the Messaging tab used to work only for Bunderstack applications. A Worker app named its provider in code with ses() or resend(), so moving from hosted mail to your own Resend account meant a code change and a deploy.

Workerstack 0.5 adds the same email() channel. Declare a channel with it and the tab offers Hosted and Resend, the connection each one needs, and a sender for each channel. Production and open previews redeploy when you switch. Agents can do the same with set_mail_provider and set_mail_sender.

To move an existing Worker app, upgrade it to Workerstack 0.5, replace ses({ from }) or resend({ from }) with email({ from }), and regenerate the blueprint. Until you do, the tab now explains that the channel picks its provider in code.

ImprovedMCP

Agents find a record instead of paging for it

Before, read_app_table could only page through a table a hundred rows at a time. To find one order among twenty thousand, an agent had to read them all, and usually ran out of room long before it got there.

Now an agent filters on any column: equal to a value or one of a list, before or after a value, containing or starting with some text, empty or not. It can sort by any column, return only the columns it needs, count the matching rows and continue with a cursor that costs the same on page two hundred as on page one. Range filters on dates take a plain ISO date, however the app stores it. list_app_tables now also shows each table's columns, primary key and indexes, so the agent knows what to filter on.

Nothing changes in what an agent may see: reading tables still needs the data:read approval, and columns named like passwords, secrets, tokens or keys stay out of every answer. They cannot be filtered or sorted on either.

FixedServers

Builds no longer fill a small server’s disk

On a server with a small disk, a few deploys in one day could fill it: Bunderhost let its own build cache grow to 10 GB however big the disk was, and trimmed it only at night. The build then failed while saving the image, with no space left on device.

The cache now keeps at most a tenth of the disk, between 1 and 10 GB. When less than 15% of the disk, or under 3 GB, is free, a build trims the cache to that size before it starts, and the deployment log says how much space it found and how much is left.

Nothing to do: builds trim from now on, and every connected server picks up the new nightly cleanup within the hour. It still never touches containers, volumes or images that are not Bunderhost's.

FixedMessaging

Disconnecting a mail domain works again

Disconnecting a domain under Messaging failed with Cannot delete … because it has tenant associations, and the domain stayed connected: its mail kept arriving in the project, even after you moved that domain's mail elsewhere. Amazon SES will not delete a domain that still belongs to a project's sending tenant, and the domain was never released from it.

Disconnect now releases the domain first, then removes it. Mail already received stays where it is.

ImprovedLanding

A new landing page, built around your agent

The landing page said what Bunderhost hosts, and left the most unusual thing about it for the fourth screen: an agent can do the whole job. The new page starts there. The headline has a switch between Dashboard and Claude Code, and the window under it shows the same project either way. On one side a session plays by itself inside the agent's own application, starting from an idea: the agent builds the app, puts it online, and carries on with what follows a launch. Each request is short and so is each answer: a link to approve, the address the app is live at, how many errors are open, how fast it is. Two buttons above the window change the application between Claude Code and Codex; the conversation stays word for word the same, because Bunderhost answers every agent the same way. On the other side is the project page exactly as the dashboard draws it.

The sections that follow answer what a visitor asks next. Which agents work, and the three ways one connects: the MCP server, the command line, and the docs served as plain text. What you can ask for. And what an agent can never do alone: creating a project, deploying, reading your users' data and changing sign-in each wait for a person, and secrets never pass through the chat.

The rest is the same product in a new layout: zero configuration, the dashboard, the choice between your own server (free) and managed hosting, and the one prompt that onboards your agent.

Nothing changes for signed-in accounts; the address / still opens your projects.

NewServers

Move an existing database to your server, or back

A database on your own server was a choice you made when you created a project. A project that already had data stayed where it was: the managed provider for most, which is exactly the data some of you wanted on a machine of your own.

Settings → Deployment target has a second choice now: Where the database is kept. Pick the server production runs on, or is moving to, and the next production deploy takes the database there. It stops writes, copies every table, checks each arrived whole, runs your migrations on the new database and switches production to it. Until production is live the application keeps answering and anything that writes fails, for as long as the copy takes: seconds for a small database, minutes for gigabytes.

If the deploy fails anywhere along the way, the new database is deleted and the old one takes writes again. After it succeeds the old database is kept, read-only, for seven days, and the one on your server is backed up daily from then on. The same choice moves a database back to the managed provider, which is also how a project leaves its server for managed infrastructure: the database first, then the target.

NewPreviews

Previews that start from seed data

A preview could start two ways: as a copy of production, or empty. A copy is real data in front of everyone who opens the preview. Empty keeps production data where it is, but leaves the reviewer signing up and clicking through forms before there is anything to look at.

Settings → Preview Deployments → A new preview starts with has a third choice: The application's seed data. The preview gets an empty database, and the application fills it from its own seed before it serves the first request: whatever its code says, written once. The same seed fills a fresh local database, so what a reviewer sees is what a developer sees.

The seed is part of the application, declared with database.seed in Bunderstack 0.33.1 or later; the choice appears once the default branch declares one. The starter that agents build new apps from has one already, with a demo user to sign in as. Production is never seeded: Bunderhost asks only previews opened for it, and nobody can set the variable that asks. A preview that is already open keeps what it started with. Agents set the same thing with previewData: "seed" in update_project.

NewPreviews

Previews on your own server

Production could already run on a server of yours, and its database could live there. Previews still ran elsewhere: every pull request was built and started on managed infrastructure, and a preview read its copy of the database over the internet. For a project that is on its own server so that its data stays there, that was the part that did not fit.

In Settings → Preview Deployments, Previews run now has a second choice: On your server. Each new preview is built on the server and started there as its own container, limited to 512 MB of memory and one CPU so it cannot starve production. With a database on the server, the preview reads its copy inside the server. The link in the pull request stays the same, and so does who may open it.

The choice is read when a preview is opened, so one that is already open stays where it is; close and reopen the pull request to move it. Closing a pull request removes the preview from the server. Previews share the machine with production, so set the number of concurrent previews to what the server has room for. Agents set the same thing with previewTarget in update_project.

FixedServers

The metrics agent installs on 1 GB servers

Connecting a small cloud server, such as a 1 GB AWS or Hetzner instance on Ubuntu 26.04, finished enrollment but never showed charts or logs: the server stayed at agent missing and its error read No space left on device. Recent Ubuntu keeps /tmp in memory, sized at half the RAM, and the agent's download did not fit there.

The agent now unpacks on disk instead. Servers that are stuck pick it up on their own within the hour, or at once when you press Verify connection.

ImprovedAgents

Agent approvals wait for you

When an agent needs a permission, it asks with an approval link in the chat. That link used to close after 15 minutes, and nothing in the dashboard said a request was waiting, so a request made while you were away was simply lost. Links now stay open for 24 hours, and every open request waits at the top of the dashboard for the organization's owners and admins, with Review.

An agent building a new app can also ask once instead of three times. The approval that lets it create and deploy the app can include turning on Google or GitHub sign-in, limited to that app's production. Previews and application users still need their own approval.

Reading which sign-in providers are on, and whether they work, is now part of read access, for connections made before today too. An agent can say what sign-in still needs before it asks you for anything.

FixedSign-in

Turning on sign-in connects it

An app receives managed sign-in with a deployment. Turning on Google or GitHub for the first time saved the setting, but the running app knew nothing about it until someone deployed again, so sign-in looked switched on and did nothing.

Now production deploys the commit it already runs, on its own, as soon as sign-in is first turned on — from the dashboard, the CLI or an agent. Sign-in works once that deployment is live. Later changes, such as adding GitHub next to Google, still apply without a deployment.

A preview, or an app deployed from uploads, cannot be shipped again this way. For those the settings card says the app receives sign-in with its next deployment, and get_project_readiness tells an agent the same.

NewServers

Move a project and its database to another server

Until now a project whose database was on a server could not leave that server. Now you choose another server in the project's Settings, or with bunx bunderhost target hel-1, and the next production deploy takes the application and its database there. The other server needs databases turned on under Fleet.

The database cannot be copied while it is being written to without losing what is written meanwhile, so it is frozen for the move: once the image is built, writes are refused, the database is copied to the new server and checked row count by row count, and production switches. The application keeps answering reads throughout. That takes about as long as a backup and a restore of the database, a few minutes for a small one.

A move that fails is put back: the new database is deleted, the old one takes writes again, and production never left. After a move that succeeds, the old database stays on the old server for seven days, and the old server can be removed once no preview still uses a database on it.

If the old server is gone, tick the box under the pending move: the next deploy then makes the database from the newest backup and does not try to reach the old server. What was written after that backup is not in it.

NewServers

Daily backups of a database on your server

A database on your own server keeps no history of itself, and a copy that lives on the same machine is lost with it. Bunderhost now backs these databases up: once a day production is copied into a bucket Bunderhost keeps for the project, and each copy stays for seven days. The newest one is kept however old it is.

The Backups card on the Data tab lists the copies. Download gives you one as a SQLite file, straight from the bucket. Restore… makes a new database on the server from it, checks that every table has the rows the backup recorded, switches production to it and redeploys; what production held is backed up first, and the replaced database stays on the server for a week. Back up now makes a copy on request, of production or of a preview, and it is what bunx bunderhost db pull now does for these databases.

A backup is made on the server while the application keeps writing, and is of one moment. The server is given a link for that one file and never the keys of the bucket. It needs free disk of about twice the database. If a day passes without a fresh backup, the server's page under Fleet shows an alert named after the project.

Daily backups are on for every database on a server. The switch is on the same card, and databaseBackups in update_project for an agent. If the server itself is lost, the project can be moved to another server from its newest backup.

NewServers

Keep the database on your own server

Until now every project's database lived in the managed provider, wherever the application ran. That does not suit data that has to stay in your hands or in a particular country. A project on your own server can now keep its database on that server.

Open the server under Fleet and press Turn on databases: Bunderhost installs one database service there in a minute or two. Then, when you create a project with that server as its production target, set Database to the server. The application reaches its database inside the machine; migrations, the Data and Users tabs and agents work as before, through an HTTPS address that only a signed token opens. A preview gets its own database on the same server, a copy or an empty one by the project's preview setting. Agents use set_server_databases and databaseProvider: "sqld" in create_project.

Two things are different from a managed database, and worth knowing before you choose it. There is no restore to a moment: Bunderhost copies the database off the server once a day instead, and production is restored from one of those copies. And the choice is made when the project is created: a database that started in the managed provider cannot be taken to a server yet, or the other way round. A project that keeps its database on a server can move to another server, and takes the database along.

NewPreviews

Previews without production data

A preview has always started as a copy of production: a branch of the production database and a copy of its files. That is the quickest way to see a change against real data, and it is still the default.

It is the wrong default when production data must not leave production, for a legal reason or simply because the people opening previews should not see it. In Settings → Preview Deployments, A new preview starts with now has a second choice: An empty database and no files. The preview gets the tables your migrations create, nothing in them, and an empty bucket.

The choice is read when a preview is created. A preview that is already open keeps what it started with; close and reopen the pull request to start it again. Agents set the same thing with previewData in update_project.

ImprovedPersonal apps

A new app starts with every feature wired

An app an agent creates without a repository starts from a Bunderstack starter. Until now it had sign-in, a database and a few tasks. Now each part of the platform is used by something on the page: Google and GitHub buttons appear on sign-in once you switch them on in Settings, attachments go to a private bucket, a project cover goes to a public one, and Email me a digest queues a job the worker sends.

Tasks can also arrive by email and by webhook. Once a domain receives mail for the app on Messaging, the project page shows an address like tasks+project_7k2@your-domain, and mail from the project's owner to it becomes a task. The signed POST /api/webhooks/tasks does the same for other systems once TASKS_WEBHOOK_SECRET is set on Env. The starter also has a newsletter form and a sign-up box that add contacts, and its public bucket is where Broadcasts puts images.

Existing apps are unchanged. The same changes are in the Bunderstack SaaS template, whose README lists where each feature lives.

NewProjects

See what a project is made of

A project's Overview now opens with a map of the environment you are looking at. On the left is everything people meet the application through: its domains, and each kind of messaging channel it declares, such as email or Telegram. The application is in the middle with the place it runs (Fly.io, Cloudflare or your own server). On the right is what it runs on: the database, file storage, realtime and background work. Each part names the provider that holds it.

Every part is a link. The database opens Data, file storage opens Storage, domains open Settings, a channel opens Messaging, and the application opens Deployments. With a preview chosen on another tab, Overview keeps it and the map shows that preview instead.

A part that needs something from you is marked: a domain still waiting for DNS, a messenger that is not connected yet, or Redis that the application needs and has no REDIS_URL for. The counts that used to sit under Declared resources (tables, buckets, jobs) are now on the parts themselves.

ImprovedDocs

Your own Dockerfile, now documented where you look

Bunderhost has long built a committed Dockerfile.bunderhost instead of its generated image, but that was written down only where operators read it. The new Application image section explains it end to end: where the file goes, what the image must do, a LibreOffice example for converting Word and PowerPoint files, and two checks that stop a deployment while the previous version keeps serving — a RUN step that runs before any migration, and a start-command self-test that runs before traffic moves.

The documentation also gains Your own servers, Running in production (logs, errors, alerts, backups, console), Application data, Sign-in and email and Command line, and Getting started opens with a map of everything the platform does. Deploying now spells out the order of a deployment: build, migrate, start beside the live version, wait for health, switch.

Agents read the same pages: each section is an MCP resource such as bunderhost://docs/application-image, and /llms-full.txt carries all of them. get_deployment and bunx bunderhost status say whether an image came from your Dockerfile.bunderhost or the generated one.

NewCLI

Roll back, read runtime logs and get the Dockerfile from the CLI

bunx bunderhost rollback lists production's recent releases. rollback --previous ships the one before the current release again, or name any of them by id; the command waits until the release reports healthy, like redeploy. Database migrations are not reversed, and on a project that deploys on push the next push deploys again.

bunx bunderhost logs --runtime reads what the running application printed, with --since 1h, --level error, --query and --env pr-12. Without --runtime, logs still prints a deployment's build output.

bunx bunderhost dockerfile prints the Dockerfile Bunderhost generates for the project, and --write puts it at the repository root as Dockerfile.bunderhost: add the system packages your app needs, commit it, and the next deployment builds it. status now also says which Dockerfile built a deployment.

FixedStorage

Files in public buckets load again

A Bunderstack bucket declared visibility: 'public' sent readers to the storage host directly, but Bunderhost's storage buckets are private, so every such file answered 403 Access Denied.

Applications now serve those files themselves at their own address under /api/files, with a day of caching on Bunderstack 0.32.1. Nothing changes in your code; the fix arrives with the next deploy of each application.

NewBroadcasts

Broadcasts: write to your contacts, or let your agent draft it

Applications on Bunderstack 0.32 can send broadcasts: one email to many contacts, chosen by tags, source and, if you like, consent. The Broadcasts tab next to Contacts has an editor with a live preview and recipient count, a test send to yourself, and send now or at a time you pick. Insert image stores a picture in the application's public bucket and puts it in the email. Counters show what was sent, delivered, opened, bounced and unsubscribed.

The application sends it through its own email channel, so the provider and sender you chose on Messaging apply; a separate channel on a subdomain like news.shop.example keeps complaints away from your sign-in mail. Every message carries a one-click unsubscribe link, and addresses that unsubscribed, bounced or complained are never written to again.

Agents can draft, edit and test a broadcast with the new tools. Sending is the one step they cannot take alone: send_broadcast returns a link, and an owner or admin approves it here after seeing the email and the recipient count. A draft changed after the request is not sent.

NewContacts

Contacts: the people your app writes to

A waitlist, a newsletter form, a lead from a landing page: until now there was nowhere to keep people who are not users. Bunderstack 0.31 adds contacts to every application. Your code adds them with app.contacts.upsert(), or turns on contacts.capture to collect them from your own forms, and every address gets a one-click unsubscribe link.

The new Contacts tab, next to Users, shows them per environment with their tags, source, consent and whether they unsubscribed. You can add one, import a CSV (addresses already on the list are merged, not duplicated), export what a filter shows, edit tags, unsubscribe someone or erase them.

Consent is something a contact can have, not a requirement. Agents can do the same with the new list_contacts and upsert_contacts tools once you approve contacts:read or contacts:write. To start, upgrade the application to Bunderstack 0.31, run db:generate and deploy.

FixedBilling

Cloudflare usage reads the 25 billion row allowance correctly

Cloudflare states each included allowance inside the name of the billing line, such as "first 25B are included". Bunderhost understood thousands and millions but not billions, so the rows-read allowance was read as 25 rows and the card showed usage far above 100%.

Billions are understood now, and the rows Durable Objects read and write are counted toward the project that used them instead of Other. Nothing was charged on the wrong figure; only the card was wrong.

ImprovedServers

Move a server to a new address without moving its apps

A cloud VM that is stopped and started often gets a new public IP. Until now a server's address could not be edited while projects deployed to it, so the only way out was to move every project off first. Now Edit address works on any server, and Verify connection finishes the move.

The verify only accepts the same machine: the server at the new address must present the host key it was first verified with. When it does, the apps on it keep running, and the DNS records Bunderhost created for them point at the new IP. Custom domains are yours to change; the verify lists every one whose A record still names the old address.

Agents can do the same over MCP: list_servers and get_server show each server's status, last error, open alerts, projects and recent failed deployments, and update_server with verify_server re-address a server.

NewMessaging

Choose where email sends from the Messaging tab

Until now an application chose its mail provider in code: ses() for hosted mail, resend() for your own Resend account. Moving from one to the other meant a code change and a deploy.

Bunderstack 0.30 adds email(), a channel that leaves the choice to Bunderhost. When an application declares one, the Messaging tab opens with Email delivery: pick Hosted to send through Amazon SES from your verified mail domains, or Resend to send through your own account. The connection each needs sits right underneath, and production and open previews redeploy when you switch. Under Senders each channel can get its own address, so a newsletter can send from a separate subdomain such as news.shop.example and keep its reputation away from your sign-in mail. Agents can do the same with the set_mail_provider and set_mail_sender tools.

Receiving is not affected by the switch: mail to your hosted domains still reaches your receiving channels when you send through Resend. Channels declared with ses() or resend() stay pinned to that provider. More providers will join the list.

ImprovedProjects

A project header you can read at a glance

The top of a project page used to take five rows before the page itself began: a breadcrumb, the name, the repository, a blueprint line and the tabs. The name appeared twice, and the mark next to it said Ready, which described the blueprint and not the application.

It is now two lines. The first is the project's name. The second answers what you usually open the page to check: whether the application is live and at which address, the repository and branch, the commit that is running and when it was released, and the blueprint with a small refresh button beside it. The address opens the application, so the separate Visit button is gone, and the commit opens that release in Deployments.

While a release is building or deploying, the status says so instead of Live. A blueprint that failed or was never read still gets its own panel, because that needs your attention.

The tabs under the header look like bookmarks again: the group you are in is outlined and open towards the page it belongs to.

ImprovedErrors

Errors say why, and skip what is not a bug

A server error such as Failed query: select … used to stop there: the query was named, the reason was not. From bunderstack 0.29.1 the stack ends with a Caused by line for the error underneath, such as a dropped database connection, and the job log lines carry it too. Issues still group by the top message, so nothing you already resolved reopens. Upgrade an app to bunderstack 0.29.1 to get it.

The browser reporter now leaves out errors that are not defects. A call your API refused with 401 or 403 because the session ended or the user may not do that, a view transition the browser skipped because the tab was hidden or the window resized, and an Android in-app browser closing its bridge no longer create issues. A real crash on the same page is still reported.

NewDocs

Documentation is public, with a page written for agents

The documentation used to sit behind sign-in, so an agent could not read a word of it before you had connected it, and neither could a person deciding whether to sign up. It is now open at /docs.

There is a new page next to it, /docs/agents, written for the agent rather than for you: which stacks Bunderhost runs, how to start an application from an empty directory, how to connect to your account, and the steps from a source tree to a live URL. It carries the same guides the agent reads after it connects, so the two cannot disagree. The same text is served as plain Markdown at /llms-full.txt, with an index at /llms.txt.

The prompt on the landing page, on an empty Projects page and under Agent access is shorter because of it. It tells the agent to read that page, where to connect, and the two things that still need your yes: creating the project and deploying to production. Secrets stay out of the chat as before.

ImprovedDashboard

One look for the landing page and the dashboard

Until now the landing page and the dashboard were two designs: warm and editorial on one side, cold blue-black with a neon accent on the other. Signing in felt like arriving at a different product. Both now use the same warm paper and ink, one ultramarine accent, the Switzer typeface and angular shapes.

Inside a project, the fourteen tabs no longer run off the edge of the screen. They are grouped by what you came to do: Overview, Deployments, Monitor (Analytics, Errors, Performance), Logs (Logs, Background, Console), Data (Data, Storage), Users (Users, Messaging) and Settings (Settings, Env). Every section keeps its address, so existing links still open the same page.

A few smaller things follow from the same pass. A status is a small square mark and a word instead of a tinted pill. Filled buttons are ink-coloured, and the accent is kept for state and links. A blueprint that is simply up to date is one fact in the project header instead of a panel.

The landing page itself went from ten sections to five and starts with what Bunderhost is: hosting for web apps.

NewProjects

A first-run checklist instead of an empty page

A new organization used to open on an empty Projects page with one button. It now opens on the shortest way to a running app.

Step one is a prompt you copy and give to your agent: Claude Code, Cursor, ChatGPT or any other MCP client. The agent connects to your account, and you approve what it may do in the browser. Secrets are entered there too, never in the chat. The project then appears on this page with its address, logs, data and users.

If you would rather do it by hand, the same page links to importing a GitHub repository or uploading a ZIP, and to adding your own server, where apps cost nothing to run here.

NewOrganization

Change member roles and transfer ownership

Every member's role in Organization → Settings → Members is now a menu. Owners and admins can switch other members between Admin and Member. Admins manage members, servers and settings, and an agent connected by an admin gets write access; an agent connected by a member can only read. The change applies at once, including to connected agents.

The owner also gets Transfer ownership in the danger zone at the bottom of the page. Pick a member and type the organization's name to confirm. They become the owner and you become an admin, so only they can give it back.

FixedDashboard

Switching organizations updates the project list

Choosing another organization in the sidebar switched the session, but a projects page that was already open kept the previous organization's list until you reloaded the page. Live updates also kept following the organization you started in.

The page now loads the new organization's projects as soon as you switch, and live deployment and status updates follow you to it. The same happens when you accept an invitation or create a new organization.

ImprovedMCP

Agents can ask for an organization you joined later

An MCP connection covers the organizations you selected when you approved it. Before, if an agent asked about any other organization, Bunderhost answered as though the sign-in had expired, and clients such as Claude Code dropped the whole connection.

Now the agent sees which of your organizations are outside the connection and can ask for one. It gives you a link; open it while signed in and choose Connect. Your other organizations stay connected, no token is shared, and deploy permissions apply only where you are an owner or admin.

Reconnecting from scratch also keeps what you chose before: the consent screen starts from your last organizations and access level. Approval links for extra permissions work whichever organization the dashboard has open.

NewBilling

See what your projects use: compute, database and storage

Until now Bunderhost could tell you about your database usage and your balance, and nothing about the machines your apps run on or the files they store. The new Usage this month card puts all three together: for every project, running machine-hours on Fly, database storage and rows on Turso, and object storage on Tigris, each priced at the provider's list price, with the organization's total and where it lands by the end of the month at the current rate.

Previews count too — a pull request's machine and bucket show up under the project that owns it. The free allowances (5 GB of database storage, 500 M rows read, 10 M rows written, 5 GB of object storage) are shared by the whole organization and shown as meters, so a project inside the allowance costs $0.

Each project's Overview has the same figures for itself, split into production and previews, and agents get them from get_project and the new get_organization_usage tool. Nothing is charged by any of this: the numbers are what the usage would cost, so you can see it before billing exists.

FixedGitHub

Connecting GitHub no longer waits for an expired upload

A project that deploys from ZIP uploads can switch to a GitHub repository from its settings. That switch refused with "Finish the active upload before connecting GitHub" whenever an upload was still labelled ready — including uploads whose 30-minute deadline had passed hours ago and that only the nightly sweep would have relabelled.

The dashboard now applies the same rule the MCP tool already did: an upload blocks the connection only while it is still before its deadline, or while the deployment it fed is still running.

If you hit that message today, try again: it goes through.

NewGitHub

Move a GitHub account to another organization

GitHub allows one installation of the Bunderhost App per account, and Bunderhost connects that installation to one organization. Until now, connecting an account that another organization had connected ended with "disconnect it there first" — and there was nothing to disconnect it with.

Now the same Connect GitHub button handles it. GitHub asks you to sign in and confirms you administer the account; Bunderhost then shows what the move changes and asks you to confirm. Projects of the previous organization keep deploying from their repositories, but only your organization can connect further repositories from that account.

Self-hosted Bunderhost needs the App's OAuth client (GITHUB_APP_OAUTH_CLIENT_ID and GITHUB_APP_OAUTH_CLIENT_SECRET) for the confirmation step; without it the page says so.

FixedDashboard

Project icons for apps on Cloudflare and your own servers

After a deployment goes live, Bunderhost reads the app's HTML once and keeps the icon it declares, so the projects list and the project page show your favicon instead of a letter. That lookup only ran for deployments on Fly.

Projects on Cloudflare and on your own servers now get the same lookup after every production deployment. If the app declares an icon, it appears within a few seconds of going live; if it does not, the dashboard keeps trying /favicon.ico on the app, as before.

Nothing to do: the next production deployment picks the icon up.

NewGitHub

Know when GitHub stops reaching your repository

A GitHub App only receives pushes for the repositories its installation covers. When an owner of the GitHub account narrows that list, or removes the App, pushes to the affected repositories stop arriving — and until now Bunderhost stopped deploying them without a word. A project could sit on a week-old build while its repository kept moving.

GitHub does report the change, and Bunderhost now acts on it. The project page shows a notice naming the repository and when access was lost, with a link to the installation's settings on GitHub. Owners and admins of the organization get an email, and paired Telegram chats a message, listing the affected projects.

The notice clears itself as soon as access is back: when the repository is added again, when the installation returns to all repositories, or on the next push, which GitHub only sends for repositories the App can reach.

FixedGitHub

Connecting GitHub says why it failed

Connecting GitHub from a project's settings sends you to GitHub and back. When Bunderhost refused the connection on the way back, every refusal looked the same: a bare Internal Server Error, with nothing to act on.

Now you land back on New project, and the page says what happened: the connect link expired, the App was installed before the link was made, or you are not an admin of the account. Each one tells you what to do next.

If you saw that error before, start again from Connect GitHub.

ImprovedBackground

Background run history moves out of your database

Until now every background run, cron minute included, stayed as a row in your application's database for a day before it was swept. That was a write, a rewrite and a delete per run on a database you pay for by the row, and it was still only a day of history.

With bunderstack 0.29 a job's row is gone the moment it succeeds, and a schedule owns one row that moves from slot to slot. What a run was is reported with the app's telemetry and stored with its logs, where the Background tab lists it: every attempt with its outcome, duration and error, for 30 days instead of one. Failed rows stay in the queue, with their payload, for a retry.

Update bunderstack to 0.29.0 in the project; nothing else changes in the application. Until then the tab shows the queue as before.

NewDashboard

Release notes when you come back

Bunderhost ships often and nothing in the product said so. Opening the dashboard after a release now shows a short dialog with every release since your last visit, each change as a one-line title with a sentence underneath.

Dismissing it is remembered on your account, so another browser or device does not show the same notes twice. The sidebar's What's new entry reopens the recent releases at any time, and All changes leads here.

NewLogs

Runtime logs for Cloudflare Workers

Until now the Logs tab of a project on the Cloudflare target said that logs were not available yet, and get_runtime_logs had nothing to give an agent. The Worker's output stayed in Cloudflare's own dashboard, which the project's owner cannot see.

Cloudflare now exports each Worker's logs to Bunderhost, which stores them next to the logs of self-hosted apps. The Logs tab, the get_runtime_logs MCP tool and the environment diagnostics read them the same way: searchable for 30 days, across deployments, with the timeline and level counts.

Nothing changes in the application. The platform picks the export up at the project's next deployment.