Open Source · MIT License

Your agent can run
its own hosting platform.

Install agentic.hosting once on any server. It becomes your platform — a place where every agent you build can deploy and operate its own projects. Blogs, APIs, web apps, databases. One agent keeps it running. The rest just ship.

Read the Agent Runbook

Bring your own Linux server · no vendor lock-in · $0/mo platform fee

agent session

# Step 1 — your webmaster agent sets up the platform (once)

$ curl -fsSL https://agentic.hosting/install.sh | bash

✓ Your platform is live. Hetzner CX33: $9.99/mo before IPv4 and VAT.

# Step 2 — now ANY of your agents can deploy projects to it

blog-agent: /ah-deploy github.com/me/my-blog blog

✓ live at https://blog.your-server.traefik.me

api-agent: /ah-deploy github.com/me/my-api backend

✓ live at https://backend.your-server.traefik.me

The problem

Standard cloud hosting wasn't built for agents.
We started over.

Every existing hosting platform was designed around a human at a keyboard — dashboards, wizards, OAuth flows. Agents hit walls humans never notice. We rebuilt from scratch so agents can operate the whole platform themselves, without a human in the loop.

JSON-only responses

No more HTML error pages.

Standard platforms return HTML 500 pages that break LLM parsers. Every response from agentic.hosting — including errors — is JSON. Errors are {"error":"..."}. Your agent always gets a parseable body.

Idempotency keys

Retries don't create duplicates.

Agents hallucinate, timeout, and retry. Authenticated write endpoints accept an Idempotency-Key — same tenant + method + path + body hash returns the cached result. Keys expire after 10 minutes. POST /v1/environments/{id}/exec is excluded.

Circuit breaker

5 crashes in 10 minutes and the loop stops.

Autonomous agents can crash-loop a service at machine speed. The built-in circuit breaker detects the pattern and pauses the service automatically — before it burns your server. Auto-recovery after 30m, 1h, 4h.

Scoped API keys

Limit what your agent can touch.

Give your agent a key scoped to one tenant. If the agent is compromised via prompt injection, it can't reach other tenants, revoke keys, or delete production databases. Blast radius is bounded by design.

The API

One API. Your agent does the rest.

Deploy services. Provision databases. Stream logs. Check health. All from a single REST API your agent already knows how to call. A lightweight Go binary orchestrates Docker and Traefik under the hood — nothing exotic, nothing proprietary.

✓ Deploy any Docker image — running in under 30 seconds
✓ Postgres & Redis on demand — connection string auto-injected, agent never sees the password
✓ Reconciler re-syncs every 30s — crashed containers restarted automatically
✓ Per-tenant isolation — each agent gets its own network, rate limit bucket, and key space

agent deploying a service

# Deploy a Docker image

$ curl -X POST https://agentic.hosting/v1/services \

-H "Authorization: Bearer $AH_KEY" \

-d '{"name":"web","image":"nginx:alpine","port":80}'

{"id":"svc_01jn...","status":"deploying"}

# Provision a database — one call

$ curl -X POST https://agentic.hosting/v1/databases \

-H "Authorization: Bearer $AH_KEY" \

-d '{"type":"postgres","name":"mydb"}'

{"status":"running","connection":"[encrypted]"}

✓ DATABASE_URL injected · service restarted

container runtime

// every container, every deploy

Runtime: runsc (gVisor — not runc)

CapDrop: ALL

ReadonlyRootfs: true

NoNewPrivileges: true

NetworkMode: tenant-isolated

PidsLimit: 256

// circuit breaker

! web: 5 crashes in 10m → circuit open

→ paused · auto-retry in 30m

Safety

Safe to let an agent deploy — even code you didn't write.

The hardest part of autonomous agents isn't making them work. It's trusting them. gVisor runs every container in a user-space kernel — agent-generated code can never reach the host. And if your agent hallucinates a crash loop, the circuit breaker catches it before you get paged at 2am.

✓ gVisor kernel — container syscalls never reach the host
✓ All Linux capabilities dropped, read-only rootfs
✓ Per-tenant networks — containers can't see each other
✓ Circuit breaker kills runaway loops automatically

Claude Code

Or just say /ah-deploy.

Install the skill once (install-skill.sh). Your agent gets six /ah-* slash commands that know how to use the API — no curl to construct, no JSON to format.

/ah-deploy

Docker image or git repo. Streams progress.

/ah-status

Full health dashboard — disk, services, circuits.

/ah-db

Provision Postgres or Redis, auto-wire to service.

/ah-logs

Stream build logs for any service by name.

Claude Code

you /ah-deploy github.com/me/my-blog blog

→ Cloning · Building · Starting...

✓ live at https://blog.65.21.x.x.traefik.me

you /ah-db blog postgres

→ Creating database...

✓ database wired up · blog restarted

you /ah-status

Disk: 24% · all systems ok

✓ blog — running · 0 crashes

✓ blog-db — running

Pricing

Your server. $0/mo platform fee. No surprises.

agentic.hosting is a single Go binary you run on your own server. A Hetzner CX33 has 4 vCPUs, 8 GB RAM, and 80 GB NVMe. Hetzner lists it at $9.99/mo or €8.49/mo before IPv4 and VAT; a primary IPv4 adds $0.60 or €0.50 per month. Service capacity depends on workload. You own the server. You own the data.

$20/mo + usage

Vercel Pro

Per-deploy charges. Bandwidth fees. Scales up when your agent loops.

$20/mo + usage

Railway

Per-minute container pricing. Dashboard required for setup.

your server

$10.59/mo + VAT

Hetzner CX33 with IPv4

4 vCPUs, 8 GB RAM, 80 GB NVMe. No per-deploy or per-seat fee. Server rate + IPv4 rate.

Your agent can runits own hosting platform.

Standard cloud hosting wasn't built for agents.We started over.