Alter Docs MCP Server

The Alter Docs MCP server is a hosted MCP server that serves Alter’s documentation and setup guidance to your coding agent. Connect it once, then say “set up Alter” — the agent reads the docs itself, classifies what you are building, walks the setup steps, writes the SDK into your codebase, and checks the verification gates.

https://mcp.alterauth.com/mcp

It is public, unauthenticated, and read-only. It holds no account credentials, never touches the vault, and cannot change anything in your account. Account actions stay with the alter CLI, which the agent runs in your own shell under your own login.

Warning

This is not a credentials server. The Alter Docs MCP server gives your agent documentation, not API access. If you want to expose credentials Alter manages to Claude Code as callable tools, that is the opposite direction — see Expose credentials to Claude Code, where you build and run your own MCP server holding an agent key.

Connect it

claude mcp add --transport http alter-onboarding https://mcp.alterauth.com/mcp

Add to .cursor/mcp.json in the project (keep any servers already there):

{
  "mcpServers": {
    "alter-onboarding": {
      "type": "http",
      "url": "https://mcp.alterauth.com/mcp"
    }
  }
}

The server speaks streamable HTTP at https://mcp.alterauth.com/mcp. Most clients take the same JSON shape:

{
  "mcpServers": {
    "alter-onboarding": {
      "type": "http",
      "url": "https://mcp.alterauth.com/mcp"
    }
  }
}

alter init writes the entry for you, merging into an existing config rather than overwriting it. A .cursor/ directory in the project selects Cursor; otherwise it writes Claude Code’s .mcp.json. Pass --client claude|cursor to choose explicitly.

npx @alter-ai/cli init

Already have the CLI installed? alter init does the same thing. See alter init for the flags.

Restart the MCP client so it picks the server up. Every route above registers the server under the same name, alter-onboarding, so they refer to the same thing — though claude mcp add records it in Claude Code’s own configuration while alter init writes the project’s config file.

The same setup — CLI install and sign-in, the MCP connection for the chosen client, and a first prompt — is shown in the developer portal itself: on the dashboard as Build with your coding agent, and behind the CLI & MCP button in the header (inside an application it adds the alter link step pinned to that application). The portal’s copy of the sign-in command carries the flags the current environment needs, so it can be pasted as-is.

Note

https://mcp.alterauth.com/mcp is the only MCP endpoint Alter operates. Alter will never ask you to add a different MCP host, and connecting one is how an attacker would put themselves between your agent and its instructions.

Use it

Ask the agent for what you are building, in your own words:

Set up Alter so a user can connect their Slack workspace and my app bot can post in a channel they approve.

The agent calls get_started, classifies the use case, and walks the flow — creating the app, minting a scoped key, configuring the provider, wiring the SDK, setting the runtime policy (including a human-in-the-loop gate when the use case needs one), and confirming a real call produced audit events. OAuth consent and other browser steps are handed back to you explicitly.

An integration that already works has its own phase. Ask for the change in your own words:

Require a human to approve every payment this agent makes through Stripe.

The agent picks the matching modification flow — add a provider or secret, add an agent, rotate a key, manage a grant, or set a runtime policy — and, for a policy, dry-runs the rule with alter policy simulate before it is saved. The Runtime policies section of the developer portal, the Wallet’s Runtime Policy tab, the alter policy CLI, and the policy_language tool below all author the same rules; the portal’s Human in the loop (HITL) type is the require_approval rule type everywhere else.

What it exposes

Your MCP client discovers the live tool set through tools/list; that runtime response is authoritative. This reference is checked against a real MCP tools/list round trip so a newly registered tool cannot remain undocumented here.

Tool	What it does
get_started	Return phase-scoped flows and a non-authoritative routing hint; the calling agent classifies the use case.
next_step	Advance to the next step of a flow. At the design step it repeats that flow’s starter ALTER_INTEGRATION.md.
sdk_integration	Return the SDK wiring to write into the codebase.
sdk_pattern	A runnable call pattern for a language.
verify_integration	A copy-pasteable recipe to prove the integration works.
troubleshoot	Map a CLI exit code or error to a remediation.
list_providers / list_operations / get_operation_schema	Browse the live provider-spec catalog and an operation’s exact contract.
policy_language	The authoritative policy-language grammar, live from the backend.
fetch_doc	Fetch any page of these docs by slug — the server bundles the entire published documentation, including the skill pages, so nothing has to be looked up elsewhere. A page reference is accepted in any spelling these docs use: a bare slug, a leading slash, a #section anchor, a full documentation URL, or an older path that now redirects. A mistyped slug is answered with the closest pages. Omit the slug to list every page.
search_docs	Keyword search across every bundled page (docs and skills): the best-matching pages with the section and snippet that matched. Use it to find the right page before fetch_doc, instead of guessing slugs or listing every page.
list_phases / list_skills	Discover what the server covers.

The full agent instructions this server serves are published as skills: Onboarding (first integration), Modification (change one that already works), and Server Integration (multi-user services).