Skip to content

The fastest, cheapest, most efficient open-source agent harness. Run more than 500 agents on a $10 VPS.

3.8 KB
Raw Blame

description: >- Put a workspace folder on the internet: deploy a site, set up a database, set environment variables, attach a domain, and roll back. icon: cloud-arrow-up

Hosting

The agent can take a folder in your workspace and make it a live site, with a database behind it if the app needs one. Ask for it in chat. The deploy is a tool call like any other, and an irreversible one goes through the approval gate.

{% hint style="info" %} Vercel is the only provider today. The design is provider-neutral, and Railway, Cloudflare and a self-hosted target are planned, but nothing else works yet. {% endhint %}

What the agent can do

Tool Effect
hosting_launch_site Deploys a workspace directory as a live site. It can also set up a database and connect it, set environment variables, and attach domains.
hosting_deployment_status Whether a build has finished.
hosting_list_deployments A site's recent deployments, newest first, with status and target.
hosting_deployment_logs A deployment's build and runtime log events.
hosting_rollback Points production back at an earlier deployment that already built.
hosting_list_sites The sites on the account.
hosting_set_env Sets environment variables on an existing site.
hosting_add_domain Attaches a custom domain.
hosting_domain_status Whether a site's domains are verified and serving.
hosting_analytics Traffic over the last N days.

A launch runs five steps in the only order that works: create the site, create the database, connect the database before the build, set the environment, then build. A framework that reads its environment at build time breaks in any other order. The call returns while the build is still running, so the agent polls hosting_deployment_status and does not block.

There is a rollback but no separate promote, because a rollback is promoting an older deployment. It refuses a deployment that never finished building. Promoting a failed build would take the site down while you try to bring it back.

What it will not do

  • Read a secret. The provider puts a managed database's connection string into the site's environment. OpenHuman learns the variable names and never their values. That is why a launch reports DATABASE_URL and not a URL.
  • Deploy something you did not name. The directory check refuses an absolute path, a .. escape and anything that is not a directory. It is the one place that decides what may leave your machine, and a deployment uploads every byte under the folder it is given.

Setting it up

[hosting]
provider = "vercel"
api_key = ""        # leave blank to read the provider's own env var
team = ""           # blank means your personal account

With api_key blank, the credential comes from TINYHOSTS_VERCEL_TOKEN, then VERCEL_TOKEN. A team account also reads TINYHOSTS_VERCEL_TEAM_ID or VERCEL_TEAM_ID.

[components] hosting is true by default and is the only on/off switch; a legacy [hosting] enabled = false is migrated once to [components] hosting = false (see Components). When it is off, or when no credential can be found, the ten tools are not registered at all. A tool that is present but cannot work is worse than a missing one, because the model keeps retrying it.

A misconfigured section is logged as an error and not skipped silently. That means an unknown provider name, or a key that is blank after trimming. Leaving api_key = "" is the supported way to defer to the provider's environment variable and is not a misconfiguration.

See also

  • Cloud deploy: hosting the OpenHuman core itself, which is a different thing.
  • Approval gate: how a deploy gets your yes.
  • Coder: building the thing before you ship it.