Release any system.
You push. RAS writes the pipeline, takes the build to everyone a stage at a time, and takes it back when it is worse.
Most teams have nobody whose job is the release. So it is a script, a wiki page and a held breath.
RAS is the release done the way a good platform engineer would do it: in stages, watched, on the record, and taken back the moment it is wrong.
One release, from a commit to everyone.
A commit is built once.
The pipeline builds an image, hands it to RAS and is done. It holds no key to any machine.
version.pushed, release.startedStaging puts it under load.
RAS drives requests at the new version and holds its numbers against the version before. On our lab: 800 requests, none failed, 1.1 ms at the 95th percentile.
stage.started, step.applied, step.passedCanary goes first.
A part of the machines runs the new version while RAS watches it against the one it replaces. With too little to judge by, it waits. It never guesses.
step.applied, step.passedThen it waits for a person, if you want one.
Approving is one click or one command, by someone the policy allows. An agent can never give it.
gate.waiting, gate.approvedProduction, a part at a time.
Ten percent, then half, then everyone, with a hold at each. Then one sentence says it is live, and how long it took.
step.applied, step.passed, release.completedA worse version is taken back.
When a check fails, RAS returns the stage to what ran before, by itself, and says why. Nobody is woken.
step.failed, stage.reverted
A service says what it is. RAS works out the rest.
ras attach reads the repository, its Dockerfiles and its Compose file, and drafts this file with the reason for every line, and what it couldn't tell. You review it and commit it. From then on a push is a release.
- No pipeline to write
- RAS proposes the pipeline too: build, check, hand over.
- Secrets are born on the machine
- A password a service needs is made where it runs. RAS never sees it.
- Jobs before, after and on a clock
- A migration before traffic, a test after it, a digest every night.
ras.yaml, and why each line is what it is
service: shop-cartThe repository, then the name Compose gives it.alias: cartWhat its neighbours in that file already call it.build: { dockerfile: src/Dockerfile, context: src }src/Dockerfile is the service's Dockerfile, so the build is the image it describes.listens: { port: 7070 }compose.yaml exposes its port 7070.CACHE_URL: { from: redis, as: "redis://{address}/0" }Compose points it at redis, so on each stage it is given that service's address.DB_PASSWORD: { secret: db-password, generate: true }Its name says it is a secret, and the value Compose falls back to is for a developer's machine, so each machine makes its own.settingscompose.yaml gives it more settings from a file that isn't read here. Write them here, the secret ones as secrets.
People and agents come in through the same door.
An agent acts as itself, over MCP, only inside a grant and always with a reason. What it does sits in the ledger beside what you do. It holds no key to any machine.
- pause a release
- object to one
- run a drill
- release on request
- attach a repository
- approve
- skip a wait
- add a machine
- change a grant
- release-reviewerpaused cart on canary
p95 doubled since the last version
- release-reviewerasked to approve productionRefused. An agent may never approve.
- a personresumed cart
one slow request, not the version
- a personapproved production
Nothing happens off the record.
Every event carries the hash of the one before it. All of RAS's state is that ledger, replayed, so there is no database to lose. We killed RAS in the middle of a release. Five seconds later it was back, and the release finished.
ras ledger pull backup.jsonl copies it to you and checks every link.
It uses what you already run, and replaces none of it.
- Machines
- Bare metal and VMs, as Docker containers or systemd units. The machine calls RAS. RAS never connects in.
- Kubernetes
- One agent inside the cluster, a namespace for each project. Tested on k3s.
- Cloudflare Workers
- RAS uploads the Worker and moves requests to it by share.
- GitLab
- Self-hosted or not. The pipeline is drafted for you. GitHub is next.
- Slack
- Told when a release needs a person, fails, or is taken back.
- Your agent
- One MCP endpoint and six written skills. Any agent that speaks MCP can use them.
One Go binary. No database. One dependency beyond the standard library.
Three commands to a first release.
On the machine that will be RAS
ras init -url https://ras.example.com \ -gitlab https://gitlab.example.com -stages staging,production
For each machine that will run your services
ras host add -target staging
In a checkout of each repository
ras attach