deploy-app
Deploy a project from a git repository on a pod: frontend built to static files, backend as a process, both behind one domain.
# Deploying an app on AgentsPodium
One pod runs both halves of a web project: the frontend build is served as
static files, the backend runs as an ordinary process next to it, and one
proxy inside the pod puts them behind a single domain with TLS. The disk
survives rebuilds. There is no split into functions and no per-service bill.
Status, 2026-09-13. Verified on live pods: a private GitLab repository and
a public GitHub monorepo, both Expo web builds inside a small pod, served
over HTTPS on their own hostname. node and static output are the tested
paths; python, go and docker runtimes are written but not yet proven, so
check GET /agents/:id/build rather than assuming.
Base URL for everything below:
`
https://agentspodium.com/api
`
Errors are always {"error":"CODE","message":"..."}.
1. Get a key once
Ask your operator for an API key from https://agentspodium.com/account
("API keys for agents"). It looks like ak_live_… and goes on every call:
`
Authorization: Bearer ak_live_…
`
An e-mail code also works (POST /auth/request, then POST /auth/verify),
but it needs a mailbox. The key does not.
2. Describe the project
Everything the pod needs to build and run the project goes in one app
object. Only repo is required; the rest has defaults or is detected.
field — meaning — default
repo — https git URL. SSH remotes are not supported — required
branch — branch to build — main
rootDir — directory the build runs in, for a monorepo — repository root
runtime — node, python, go, static, docker — detected from the repo
buildCommand — how to build — default for the detected runtime
outputDir — static output, relative to rootDir — none
startCommand — backend process — none
appPort — port the backend listens on inside the pod — 3000
apiPrefix — paths routed to the backend; the rest is served static — /api
env — environment variables for the app — {}
repoToken — token for a private repository — none
repoUser — username paired with the token — what the host expects
Set outputDir, startCommand, or both. A pod with neither serves nothing,
and the API says so rather than creating it.
3. Create it
engine is app, and the smallest plan is small: the build runs inside
the pod, so the plan is also the build machine, and npm ci on 1 GB does not
finish.
A real example — an Expo web app in a monorepo, static only, private repo on
a self-hosted Forgejo:
`bash
curl -s -X POST https://agentspodium.com/api/agents \
-H "Authorization: Bearer $TOKEN" -H "Content-Type: application/json" -d '{
"engine": "app",
"tier": "small",
"personaId": "personal-assistant",
"name": "hq-app",
"app": {
"repo": "https://git.defispace.com/hq/app",
"branch": "main",
"rootDir": "app",
"buildCommand": "npm ci && npx expo export -p web",
"outputDir": "dist",
"repoToken": "…"
}
}'
`
A fullstack example — React frontend and a Node API in the same repository:
`json
{
"engine": "app", "tier": "small", "personaId": "personal-assistant",
"app": {
"repo": "https://github.com/owner/project",
"buildCommand": "npm ci && npm run build",
"outputDir": "dist",
"startCommand": "node server/index.js",
"appPort": 3000,
"apiPrefix": "/api",
"env": { "NODE_ENV": "production" }
}
}
`
The reply is {"agent":{…}}. Keep agent.id. The pod answers at
agent.endpointUrl; agent.app shows the configuration as stored, and
agent.appRepoTokenEnc is "set" or null — the token itself never comes
back.
4. Watch the build, not the pod
`bash
curl -s https://agentspodium.com/api/agents/$ID/build -H "Authorization: Bearer $TOKEN"
# {"build":{"state":"building|ok|failed|unknown","commit":"abc1234",
# "startedAt":"…","finishedAt":"…","logTail":"last lines"}}
`
A pod can be alive and still serving the previous version while a new build
fails, so pod health is not build health. logTail carries the reason.
Poll every five seconds; a first build with a cold cache takes minutes.
5. Your own domain
Point an A record at 138.199.136.33, then send domain on create or set it
later on the agent page. GET /api/domains/check?domain=<name>&https=1 says
whether DNS points here and whether HTTPS already answers. The certificate is
issued automatically once the record resolves. Names under agentspodium.com
are refused: every pod already gets one.
6. Ship a new version of an app that already exists
This is the common case: the project is already deployed, you pushed a commit,
now make the pod serve it. Two things are needed and nothing else — the API
key, and the id of the instance.
If you do not have the id, list what the key can see:
`bash
curl -s https://agentspodium.com/api/agents -H "Authorization: Bearer $TOKEN"
# {"agents":[{"id":"agt_…","name":"hq-app","engine":"app","status":"running",
# "endpointUrl":"https://…","app":{"repo":"…","branch":"main",…}}]}
`
Match on engine: "app" and the app.repo you just pushed to. Then:
`bash
curl -s -X POST https://agentspodium.com/api/agents/$ID/rebuild -H "Authorization: Bearer $TOKEN"
curl -s https://agentspodium.com/api/agents/$ID/build -H "Authorization: Bearer $TOKEN"
`
rebuild fetches the latest commit on the configured branch, runs the build
again and reloads the pod's routing. The disk is kept, so node_modules and
framework caches survive and the second build is much faster than the first.
Poll /build until state is ok or failed; on failed, logTail says
why, and the site keeps serving the previous version in the meantime.
Nothing else is required. The token, the branch and the build command are
already stored on the instance.
7. Change the configuration, then rebuild
`bash
# point at another branch, fix the build command, replace the token
curl -s -X PATCH https://agentspodium.com/api/agents/$ID/app \
-H "Authorization: Bearer $TOKEN" -H "Content-Type: application/json" \
-d '{"branch":"release","buildCommand":"npm ci && npm run build:prod"}'
# apply it: rebuild from the latest commit, keeping the disk
curl -s -X POST https://agentspodium.com/api/agents/$ID/rebuild -H "Authorization: Bearer $TOKEN"
`
The PATCH answers "applied":"on the next rebuild" because the pod reads the
configuration when it starts. Nothing changes until you rebuild.
8. Private repositories
Send repoToken with a token that can read the repository. It is encrypted
at rest, handed to the pod at deploy time, and never returned.
repoUser you can leave out. Forges disagree about what goes next to a token
— GitHub takes x-access-token, GitLab oauth2, Bitbucket x-token-auth,
Gitea and Forgejo the token alone — and a self-hosted GitLab is
indistinguishable from a self-hosted Gitea by its URL. The pod tries the forms
in turn and keeps the one the server accepts; the build log names it. Send
repoUser only to skip that.
Know what you are agreeing to: the token lives in the pod's environment, and
the code you deploy runs in that pod. It is your own repository token, but
anything you deploy can read it. Use a token scoped to that one repository.
9. Instead of polling
Send webhookUrl on create and the platform POSTs signed events instead:
agent.running, agent.stopped, agent.failed, agent.deleted,
payment.confirmed, deletion.warning. The signing secret comes back once
as webhookSecret. See https://hosting.defispace.com/docs/webhooks.md
10. What goes wrong
Engine "app" needs app.repo—engine: "app"without theapp
object. The two travel together.
App requires at least the small plan—tier: "tiny". The build runs
in the pod; 1 GB is not enough to install a frontend's dependencies.
outputDir must be a path inside the repository— a leading/or a
.. segment. Paths are relative to rootDir.
Set outputDir, startCommand, or both— nothing to serve and nothing
to run. One of them has to be there.
Repository URL must use https—git@host:owner/repo.gitis not
accepted. Use the https URL and a token for private repositories.
- Build
failedwith an emptylogTail— the build produced no output
before dying, usually out of memory. Move to a bigger plan and rebuild.
Could not fetch the repository— for a private one, the token cannot
read it, or it has expired. The log shows the attempt with the token
replaced by ***; git prints the URL it tried on an auth failure, so it is
masked on the way out rather than trusted not to appear.
- Pod answers, site is blank —
outputDirpoints at a directory the
build did not produce. Check the build log for where the output went.
11. Everything else
An app instance is an instance like any other: GET /agents/:id for status,
/liveness, /usage, /term for when it expires, pause, resume, delete,
export, and the payment endpoints. See
https://hosting.defispace.com/docs/quickstart.md