Changelog
Every release’s notes, newest first. To move an installation to the latest release, see Upgrade. The source and the notes are in backmesh/backmesh.
backmesh v0.4.3
Upgrade with backmesh upgrade; there’s no need to uninstall.
Changes
- GPU runners.
backmesh-gpu-t4(spot) andbackmesh-gpu-od-t4(on demand) give a job one NVIDIA T4 and 4 vCPUs on AWS or Azure;nvidia-smianddocker run --gpus allwork in the job. GCP follows once anycloud supports T4 there. Install and status show your GPU quota, which new cloud accounts often have at 0, and how to raise it. - A re-offered job keeps its runner. When GitHub drops a job no runner took within 5 minutes and offers it again, the listener no longer stops the runner that is still starting for it.
backmesh uninstall --keep-appremoves the installation but leaves backmesh’s GitHub App installed in the org.
backmesh v0.4.2
Upgrade with backmesh upgrade; there’s no need to uninstall.
Changes
- A cheaper control plane on AWS. A new install creates the control plane’s two disks at gp3’s included speed (3,000 IOPS and 125 MiB/s), so it costs about $86 a month, as
installsays, instead of about $171. Install lowers the API VM’s disk through yourawsCLI, which needsec2:DescribeInstances,ec2:DescribeVolumesandec2:ModifyVolume; if it can’t, install says so and still finishes. - Clearer sign-in for orgs that restrict OAuth apps. The error now says what to do: an org owner clicks Grant on GitHub’s sign-in page, and anyone else clicks Request and asks an owner to approve it.
backmesh v0.4.1
Upgrade with backmesh upgrade; there’s no need to uninstall.
Changes
- Several installations per cloud account. A new installation names each cloud connection for its control plane, such as
bm-aws-k4nktn, so two installations in one AWS account, GCP project or Azure subscription no longer collide, and a reinstall afterbackmesh uninstallworks. Existing installations keepbm-aws,bm-gcpandbm-azure. uninstalldeletes the identities install generated. Where install generated a cloud identity for a connection,uninstalldeletes it, with its keys and grants, through your cloud CLI once the control plane is gone. If your CLI can’t, it prints the commands that do.- Reading an org’s ID uses your GitHub token. Install, upgrade,
statusandrotate-keysendGITHUB_TOKEN, elseGH_TOKEN, when reading an org’s ID, so a GitHub-hosted runner’s shared IP no longer hits GitHub’s anonymous rate limit, and a rate-limited read says so. The read needs no scope, and if GitHub rejects the token, it retries without one. - Several orgs share one control plane. Install and upgrade give each org’s listener half a vCPU of
bm-control’s Node, so a second org installed on 0.4.1 finds room, andbackmesh upgradesets the share on listeners created before. If a second org joined before 0.4.1 and its listener stayed Pending,backmesh upgraderefuses because that listener doesn’t answer. Runbackmesh uninstall --org <org>for that org, thenbackmesh upgrade, thenbackmesh installfor it on 0.4.1.
backmesh v0.4.0
Upgrade with backmesh upgrade; there’s no need to uninstall.
Changes
- backmesh moves to the
backmeshGitHub org. The source is at backmesh/backmesh, the release binaries at backmesh/backmesh-releases, and the Homebrew formula inbackmesh/tap. - backmesh is open source. The source is under the Apache License 2.0, and the cache action stays under the MIT License. The terms now cover only the service: backmesh.com and
api.backmesh.com. - New image paths. The listener and runner images are now
ghcr.io/backmesh/backmesh-listenerandghcr.io/backmesh/backmesh-runner. Installations on v0.3.x keep pullingghcr.io/anycloud-sh/backmesh-*, which stay published and are never deleted.backmesh upgrademoves the listener and your runners to the new images; new runners start from them once it finishes. - The CLI installs from the new org.
curl -fsSL https://backmesh.com/install.sh | shinstalls this release frombackmesh/backmesh-releases. The Homebrew formula moved to a new tap: runbrew uninstall backmesh && brew install backmesh/tap/backmesh. - Pull requests from forks read the cache only. In any repository, a fork’s pull request restores its base branch’s and the default branch’s entries and saves none, so its reruns start without what an earlier run of it saved.
- Groundwork for public repositories. Runners for public repositories are off by default and not yet supported: only private repositories can use backmesh until runners get a network of their own.
backmesh v0.3.4
Upgrade with backmesh upgrade; there’s no need to uninstall. With AWS or Azure runners, upgrade asks you to confirm the new VM types: run backmesh upgrade --plan, then backmesh upgrade --plan-id <id>.
Changes
installfinds your org. Without--org, it offers the org of the repo you’re in, or lists the orgs you’re an admin of to pick from. Without a terminal it prints the exact command to run. Your AWS CLI no longer needs a default region.- Runners use a fixed set of US regions. backmesh uses every US region on each cloud that offers all its VM types, and keeps the ones your account allows, naming any it skips.
--runnersnow takes clouds only, such as--runners aws;aws:us-east-2is no longer accepted. - Faster VM types first. Each label now tries m8i on AWS, n2 on GCP and D4ads_v5 on Azure first; m6i is no longer used.
- Cues land more often. A cued runner is timed for a slower-than-typical boot and waits up to 3 minutes for a late push, a re-cue keeps or replaces it cleanly, and cues work right after the listener restarts or upgrades.
backmesh v0.3.3
Upgrade with backmesh upgrade; there’s no need to uninstall. Install the CLI with curl -fsSL https://backmesh.com/install.sh | sh.
Changes
- Caches in your own cloud.
backmesh installandupgradecreate a cache bucket in each runner region. Swapactions/cacheforbackmesh/cache@v1: same inputs, restores fetch in parallel, and saves upload after the job.--cache-capsets each repository’s limit, 10 GB by default. - Automatic caches. With no workflow change, a job restores the toolchains and npm, Yarn or pnpm caches it used last time. Folders a workflow already caches are left alone. Set
BACKMESH_AUTO_CACHE: falseto turn it off. docker kill $(docker ps -q)no longer kills the runner. The runner’s own container is hidden fromdocker psand skipped bykill,stopandrm.- Closer to
ubuntu-latest.apt-get installworks withoutapt-get updatefirst; the tool cache is writable;LANG,ImageOSand/usr/gamesmatch;/dev/shmis half the runner’s memory. See Coming from ubuntu-latest. - Sign-in is GitHub’s device flow only.
installprints a code for github.com/login/device; the connect pages are gone and--deviceis no longer an option. - Faster starts. The runner image uses zstd layers, runners start before the listener’s housekeeping, and a launch that the cloud refuses moves to another region after 75 s instead of 3 minutes. Pins anycloud 0.1.77, which pulls the runner image with parallel unpacking.
- Cue results in
backmesh status. Start waits are split by what started the runner: a cue, a prediction or the job itself. - Every reclaimed run gets a note. A run that lost a spot VM gets a
backmesh: reclaimscheck run explaining the rerun.
backmesh v0.3.2
Upgrade with backmesh upgrade; there’s no need to uninstall.
Changes
- Free, with no billing. backmesh’s core features are free, with no credit card and no cap on jobs. Premium adds Slack monitoring, logs and runtime history; talk to us. Plans, trials and subscription checks are gone:
backmesh statusno longer shows a plan, andbackmesh upgradeno longer checks a subscription. - Quota waits in
backmesh status. Each label gets one line for the last 24 hours, with typical and slowest start times. When quota made runners wait, two more lines name the cloud, the regions and the exact quota to raise. upgrade --runnersremoves the clouds it leaves out. A cloud missing from--runnersis drained and removed, as aninstallre-run does.upgrade --planshows the removal first, andupgradewithout--runnersremoves nothing.backmesh cuein fresh sandboxes.cuetakes its sign-in fromBACKMESH_CUE_TOKEN, so an agent that can’t complete a GitHub sign-in can still give a heads-up.- Azure runners get a stop signal. On anycloud 0.1.76, a terminated Azure runner’s processes get SIGTERM first, as on AWS and GCP.
backmesh v0.3.1
Breaking: backmesh upgrade --device and backmesh rotate-key --device are gone, since a GitHub device code is now their default. Pass --connect for the backmesh.com/connect link instead.
Upgrade with backmesh upgrade; there’s no need to uninstall.
Changes
- Device codes for later sign-ins. Once backmesh’s App is on your org,
upgrade,rotate-key,cue --sign-in, aninstallre-run, andstatus,logsoruninstallreplacing an expired sign-in print a code to enter athttps://github.com/login/device.--connectprints thebackmesh.com/connect/<code>link instead. A first install still prints the link. - One line per setup problem. Before it reads regions or quotas,
backmesh installchecks that each cloud CLI you pick is on yourPATHand signed in, thatawshas a region,gclouda project andaza subscription, and prints one line with the fix for each that isn’t.--verboseadds the cloud CLI’s own error. - OAuth app restrictions named. When your org’s OAuth app access restrictions block backmesh’s sign-in, the error now names the org setting an owner changes, instead of reporting an outage.
backmesh v0.3.0
Historical: the paid plans below were removed in v0.3.2.
Reinstall required: run backmesh uninstall with your current release, then backmesh install with this one. backmesh upgrade refuses installations from before v0.3.0.
Changes
- Each job gets its own VM. Every GitHub Actions job runs on a fresh spot or on-demand VM in your cloud account, deleted when the job ends. Nothing is shared between jobs, so there are no pools to size or isolate.
- Runners in several regions. Choose where things run with
--control-plane <cloud>:<region>and--runners <cloud>[:<region>,...], on AWS, GCP and Azure. Azure runners aren’t yet proven on our own CI; that proof follows in a patch release. A bare--runners awsuses that cloud’s default regions, such as AWS’sus-regions near a US control plane;--runners aws:us-east-2,us-west-2keeps jobs in exactly those. Each job’s VM launches in the cheapest region with room and tries another when the cloud refuses for capacity or quota. Labels stay the same. - Slow starts move. A spot runner whose launches the cloud keeps refusing gets a replacement in another runner region after 3 minutes. A region whose quota can’t fit one runner is skipped until you raise it, and
backmesh statusnames the quota. - Runners start ahead of your jobs. backmesh learns each workflow’s shape from its past runs, inside your account, and starts runners shortly before later jobs queue, so most of them don’t wait for a VM. Unused runners are removed after 60 seconds.
backmesh cue --in <seconds>. When you, or your coding agent, start the checks you run before a push, runbackmesh cue --in 120with the seconds until you expect to push, 30 to 600, and the run’s first jobs find runners waiting.backmesh cue --canceldrops it. The repo and branch are encrypted to your listener; backmesh.com sees only your org, your GitHub user and the time. If backmesh.com is down, CI starts cold and nothing fails.- Sign in with a connect link.
install,upgradeandrotate-keyprint ahttps://backmesh.com/connect/<code>link to open on any device and sign in with GitHub. During install, the org owner installs the backmesh App on the same page.--deviceuses a GitHub device code instead. - See the plan first.
backmesh install --planandbackmesh upgrade --planprint what would change and what it costs, and a plan ID;--yes --plan-id <id>applies exactly that plan without a terminal. - Paid plans with a 15-day trial. Every plan starts with a 15-day free trial; checkout takes a card, charged on day 16 unless you cancel. There’s no free tier, and the free non-commercial use is gone.
upgradeis blocked while an org has no plan or an unpaid bill; jobs never stop over billing. backmesh rotate-key --org <org>. Replaces an org’s install key: it signs in an org admin, gives the listener a new key and only then revokes the old one. The billing page no longer lists or revokes install keys.- Docker and KVM without privileged mode. Jobs use their VM’s own Docker, so a port a job’s container publishes is on
localhost, as on GitHub’s runners.backmesh-kvm-*labels give/dev/kvmon AWS, GCP and Azure. - No Docker or saved cloud keys to install.
backmesh installuses your signed-inaws,gcloudorazCLI, and connects each cloud with a keyless role.upgradeandstatusneed no cloud CLI. - A fresh runner group per install, so a reinstall never reuses a deleted group’s name.
- Telemetry now includes each runner’s VM type and boot time. The telemetry page lists every field.
backmesh v0.2.1
Upgrade with backmesh upgrade; there’s no need to uninstall.
Changes
- KVM labels. Jobs that start microVMs, emulators or other virtual machines can use
backmesh-kvm-2tobackmesh-kvm-16and get/dev/kvm. Each cloud gets two more pools,bm-<cloud>-kvm-spotandbm-<cloud>-kvm-od, whose VM types support nested virtualization:m7i.4xlargeandm8i.4xlargeon AWS,n2-standard-16andc2-standard-16on GCP, andStandard_D16as_v5andStandard_D16s_v5on Azure. Like the other pools, they idle at zero VMs. A job reclaimed on spot reruns on demand, andbackmesh-kvm-od-<n>always runs on demand. With several clouds the labels name the cloud, such asbackmesh-aws-kvm-8. A region that offers none of these types gets no KVM labels. The KVM pools share their cloud’s spot and on-demand vCPU quotas with the other pools.backmesh upgradecreates them on an existing installation. Every job on a KVM VM can open/dev/kvm. - Uninstall exits cleanly. A full
backmesh uninstallthat deleted the control plane no longer exits 1 with anapi useerror, so its final report of leftovers prints. When the recorded control plane doesn’t answer, uninstall goes on to the unreachable-control-plane path instead of failing first. - Usage events name the cloud. Usage events now name the cloud of on-demand and KVM jobs, which v0.2.0 reported as unknown.
backmesh v0.2.0
Historical: the paid plans below were removed in v0.3.2.
Action required: uninstall, then install again
This release keeps no compatibility with 0.1.x installations. Run backmesh uninstall with your current release, then backmesh install with this one. What changed:
- One backmesh GitHub App. backmesh no longer creates a private GitHub App in your org. An org owner installs the public
backmeshApp with one click duringbackmesh install. The App’s private key stays with backmesh and never enters your accounts. Your listener gets hour-long tokens for your own org fromapi.backmesh.com. - User-namespace runner pools. Runners no longer run privileged. Each runner pool is a user-namespace pool, so a job is no longer root on its VM.
Your cloud credentials, code, secrets, caches, logs and runners stay in your accounts.
Changes
- One-click App install.
installsigns in an org admin with a device code, prints the App’s install link until an org owner installs it, and only then creates anything billed.upgrade --app-keyand the App key-age warning are gone, because there’s no App key in your accounts to rotate. - Several orgs on one control plane.
backmesh install --org <second>against an existing installation adds that org’s listener and shares the runner pools and the 500-runner cap with the other orgs, first come, first served. Every org must have the same owner, confirmed at the prompt or with--same-owner.statusshows each org,upgradeupdates every org, anduninstall --orgremoves one org. - One uninstall path.
backmesh uninstallalways does the same thing, and--yesskips its confirmations:- For each org, it signs in an org admin with a device code, which CI can enter from the job log. It refuses while another backmesh installation in the org has used the App in the last hour.
- It stops the org’s listener and removes backmesh’s runners, scale sets and runner group from GitHub.
- It removes the App from the org, unless another backmesh installation there has used it in the last hour.
- Once no org is left on the control plane, it deletes the cloud side if the control plane still answers. If the control plane is gone, it says so and lists what may be left.
- Retry on demand after a spot reclaim. Each cloud gets an on-demand pool that idles at zero VMs. When a spot VM is reclaimed, backmesh reruns the run’s failed jobs on demand once the run finishes, if only reclaims failed it, up to twice per run. A breaker pauses spot on a pool after repeated reclaims. To keep a job’s reclaim from rerunning its run, add
backmesh-noretryto itsruns-on. Newbackmesh-od-<n>labels (backmesh-<cloud>-od-<n>with several clouds) always run on demand. - Runner pools idle at zero VMs. Spot and on-demand runner pools scale to zero VMs when no jobs run, so an idle installation pays only for its control plane: about $86 a month on AWS, $74 on GCP and $111 on Azure. The first job after an idle gap waits a few minutes for a VM to start.
backmesh logs. Shows the listener’s log, or one runner’s log by its runner name.- Self-contained CLI. You install only
backmesh. It fetches and checks the anycloud release it pins and handles the GitHub sign-in itself, and its messages no longer ask you to run anycloud commands. - Plans. Starter (50,000 jobs per rolling 30 days, $40 a month or $350 a year), Growth (200,000; $125 or $1,050), Scale (500,000; $250 or $2,100) and Unlimited ($500 or $4,200), each with a 15-day free trial. Subscriptions open soon. Going over a tier only warns, and jobs never stop over billing. Non-commercial use (open source, non-profits, education, personal projects) is free once an org admin declares it on the billing page, with a public link back to backmesh.com.
- Revoke an install key. On the billing page, any admin of the key’s org can revoke the key. This is separate from disowning an install, which only stops counting its jobs.