When to use ohmyho.st
Use ohmyho.st when the customer wants to develop or deploy an app using its managed versions or selected GitHub source; inspect a project’s deployment, database or DNS/mail readiness; manage usage and budgets; or request an encrypted SQL export. Use the transport and operations available in this client. Your agent follows the get-started Skill for installation, sign-in and workspace selection before its first deployment (what it covers). Ask which account is signed in to read its current email and scoped workspace IDs and names throughidentity_get, on either transport. A user API token identifies its owner; machine-to-machine credentials have a null email. Current account details.
Start with identity_get, then select the customer’s project with the available list tool: remote project_list or local projects_list. Read project_context_get and source_get before acting. Use include_binding: true on source_get before using connection/generation guards. Default GitHub reads preserve their existing status fields; managed reads always include the full binding. The server’s current source binding is authoritative; an agent’s location or copied Git remote does not replace it. A new app developed entirely in chat starts with managed versions, while an explicitly selected GitHub repository uses GitHub. Configure only the capabilities the application needs and keep its chosen authentication provider. Use the matching task Skill for the workflow.
For a new project, ask whether Dev should be protected or public before project_create; protected is the default. project_dev_share_link_get returns the Owner’s reusable protected link. Local clients expose project_dev_access_mode_set to change the choice, while project_dev_share_link_rotate and project_dev_share_link_revoke block old links and browser sessions on their next request. Discover the connected client’s actual tools before requesting those changes. Dev access details.
When the client’s tools support the requested custom domain before the first Prod deployment, use domain_paid_plan and domain_paid_apply to reserve it, then domain_cloudflare_authorize and domain_cloudflare_status for scoped consent. awaiting_deployment means no DNS or certificate provisioning has started. Deploy and verify the Prod hosting address, then repeat domain_paid_apply with the original hostname and idempotency key. Consent and deployment alone do not activate it. Domain setup.
Connected chat with remote OAuth
For a client that supports authenticated remote MCP, connect:identity_get and connection_grants in the background. If you already described an application or task, it continues that request. Otherwise it asks only “What would you like to host?” in your language—for example, “Was möchtest du hosten?” in German. Routine checks need no permission table, internal IDs, empty fields or grant timestamps.
New-app consent is explicit: allow_create: true permits creation even when the approved project list is empty. An empty list without that consent does not authorize creation. allow_publish_created_projects covers only applications created under that consent; an existing application needs its own publication permission. Every operation still checks current account and project authorization.
Editing and Dev previews are separate from publication. The agent requests a publication handoff only when you have asked to publish and the relevant permission is missing. If workspace setup is needed, workspace_setup returns the account portal; the agent verifies identity again after you finish.
When project access is missing, the agent uses connection_request. Open its Review connection card when your client displays it; otherwise open the returned private portal link. Use the browser profile signed in to the ohmyho.st account connected to this client. Check the account, workspace, selected applications and requested access before confirming. The card opens the consent page; it does not approve access.
When the client provides the actual current chat address, the portal can return you to that conversation after approval. The agent may supply return_url only when it knows that exact address. Never construct it from anonymized session IDs or session metadata. Without a conversation address, a recognized client’s home page is the fallback; with an unrecognized client, return to your chat manually. Returning to chat does not confirm project access: the agent checks connection_status and connection_grants in the background before continuing.
Project-access portal links are valid for one hour. Portal links for secret input and GitHub setup last ten minutes. If a link expires, ask your agent for a fresh one.
The client manages its OAuth credential. This route needs no copied platform token or local device-code login. A plugin installation or completed browser page is not the access check: verify an authenticated tool call and the actual project grant. Discover this connected server’s tools rather than assuming the local catalog applies to it. Private application values use the portal flow in Runtime secrets.
For a complete file tree, the remote tools are source_upload_prepare, source_blob_put, source_upload, source_upload_commit and source_switch. They use actual file contents available to the client; a local directory path gives the chat no filesystem access. Blob uploads return server-computed hash/length receipts. Retain those exact receipts for the complete manifest, and require a succeeded operation plus a committed upload before using its saved SHA. Saving or switching source starts no deployment. Follow the full snapshot and switch flow.
MCP file reads return valid UTF-8 as content_text by default, or base64 when the file cannot be decoded. Remote source_files accepts path and format: "text" | "base64"; local source_file provides the corresponding file read. source_changes accepts exactly one content_text or content_base64 per change, with null base64 for deletion. Follow the returned schemas and preserve the file’s executable flag and source revision guards.
Install and sign in locally
Requires Node.js 22 or newer.@amerged/ohmyhost-cli and @amerged/ohmyhost-mcp. Both sources install the ohmyhost and ohmyhost-mcp commands, so keep one source per computer and upgrade from the one already installed (npm ls --global names it). To switch, first remove the other pair: npm uninstall --global @ohmyhost/product-cli @ohmyhost/mcp for the archives, npm uninstall --global @amerged/ohmyhost-cli @amerged/ohmyhost-mcp for npm.
Release manifest and checksums · Current release
The local stdio MCP server uses OHMYHOST_TOKEN when that variable is set in the server process, and otherwise the logins saved by ohmyhost login. With several, pass profile_name per call (profile_list shows names) or bind the server to OHMYHOST_PROFILE=NAME; another requested profile fails profile_context_mismatch. A login added later works on the next call without restarting. A token on its own is enough — no login and no browser. Put it in the same env block as OHMYHOST_ENVIRONMENT; the server reads no .env file and never calls the CLI.
With a token, every tool works except organization_create, organization_list, organization_use, token_create, tokens_list and token_revoke. Those need an interactive login: run the matching ohmyhost command in a shell without OHMYHOST_TOKEN.
After installing or upgrading the clients, restart the MCP server. A running server keeps the tool list it started with, so newly published tools stay invisible until it restarts.
Keep existing servers, models and permission settings. If a server named ohmyho is already configured, inspect it before changing that entry. An old npx configuration pinning a versioned release URL stops when that archive is superseded; use the installed ohmyhost-mcp command. Preserve harness approval policy; rejected mutations are not platform failures, and switching transport cannot bypass approval.
The server’s source is published with every release at amerged-org/ohmyhost-mcp, every file of @amerged/ohmyhost-mcp except its built dist bundle and README, byte for byte. release.json records full-package npm integrity and the SHA-256 of omitted files. For help, ask your agent: it reports through feedback_submit and follows up with feedback_status (Support).
Codex
identity_get and list the available Skills.
Claude Code
/mcp in Claude Code and confirm the connection. Ask it to call identity_get.
Cursor
Merge this server into your existing~/.cursor/mcp.json, or .cursor/mcp.json for this project:
agent mcp list-tools ohmyho. Use your installed CLI’s help if its executable is named differently.
Hermes
OpenClaw
For versions with the native outbound MCP registry:openclaw mcp add --help to confirm support. mcp serve exposes OpenClaw itself and is a different feature. Preserve the current approval policy; a saved registry entry alone does not prove that the running agent has loaded it.
Confirm the connection
Ask the agent: “Read my ohmyho.st identity and list my projects.” It should callidentity_get, then remote project_list or local projects_list, and read project_context_get for the project you select. If a tool is missing, compare the installed client with the current release and reconnect after updating.
These harness commands and the downloadable MCP configuration set up the local ohmyhost-mcp stdio server, with its CLI/device-login profiles and private automation-token support. Its published catalog describes that local client. The remote product endpoint above is a separate OAuth connection. Keep the transport your client supports and its existing permission policy. Mintlify’s search MCP answers documentation questions; it cannot deploy or access your account.
Task Skills · Actual tool catalogue · CLI