Connect your AI.
Choose your AI, add Ahel, and sign in.
Connect
Choose your AI
Connect Ahel to Claude
Add Ahel once, then sign in to authorize access.
First-time setup
- In Claude, open Settings, then Connectors, and choose Add custom connector.
- Name it ahel and paste your connector address as the server URL:
https://mcp.ahel.ai - Choose Connect, then approve the Ahel sign-in.
Claude currently requires this initial setup. Opening it here does not create a connection.
Using Claude Code?
- Run this once in any terminal; it adds Ahel for every project:
claude mcp add --transport http --scope user ahel 'https://mcp.ahel.ai' - In Claude Code, open /mcp, select ahel, and approve the browser sign-in.
Secure browser sign-in. No API keys to paste.
Part of a team on Ahel? This address connects your AI to the workspace you were invited to. In more than one? Connect your AI on ahel.ai has an address for each.
Technical details and questions
The protocol
How does MCP work?
The Model Context Protocol gives your AI one standard way to call tools. A connector is an address that answers those calls. Most connectors serve one product; this one serves your account, so the same address means something different for every person who signs in: their apps, their skills, their keys, sealed in the vault.
The tools
What the connector serves today
These are the 15 tools to use, and a client will also list 5 older names for the same jobs, still served so existing connectors keep working. That is the whole list, and it never changes: the hosted MCP servers in your workspace and the sandboxed local servers are served through this same connector, reached through tools and use rather than added to the list. Enabled skills are served too, as MCP prompts and through skill_read.
| Tool | What it does |
|---|---|
| explore | Ask for what you want your AI to do. Answers with ONE offer, the app, skill or knowledge that can do it now, and quiet alternatives. |
| install | Install one app, server or skill by id. It is added and made available to your AI clients in a single call; there is no separate enable step. |
| installed | What your account has installed, and whether each one is on, off, needs setup, or is unavailable. |
| switch | Turn an installed app, server or skill on or off. Off keeps it installed but removes its tools and prompts from your AI clients. |
| status | Per app, server or skill: its state and the names of any keys still missing. Status only: values never leave. |
| tools | The tools one installed server offers, with their schemas. This is how a server's tools are read without the connector's own list ever changing. |
| use | Run one tool of an installed server, key attached server-side. The server's text and data come back as it sent them; a server's own card is not passed through yet. |
| list_actions | What your AI can do on connected apps with your own key: ranked actions for an ask, or one service's full list. |
| run_action | Run one action on your own account of that app. Ahel injects the stored key; no key ever appears in chat. |
| raw_request | One HTTP request to a connected app's own API, for calls no listed action covers. Pinned to that app's domain. |
| run_script | A short Node script in a locked sandbox pod that holds no key; it acts through the same server-side injection. |
| skill_read | The full instructions of any installed skill that is on. The same content also arrives as an MCP prompt, which Claude Code surfaces as a slash command. |
| search | The ChatGPT connector contract: search the catalog, get id, title and url per result. |
| fetch | The ChatGPT connector contract: read one catalog record in full by id. |
| get_started | Where to begin: what Ahel is, what your workspace already has, and one thing to try first that needs no sign-in. Reads only; changes nothing. |
The boundary
What it deliberately does not do
Your AI can
- Search the whole catalog from inside a conversation.
- Add servers to your workspace and switch them on or off; that only ever changes your own workspace.
- Call the tools of every server you serve, keys attached server-side.
It never can
- Read, write, or reveal a stored key. No tool on the connector touches a key, and a key never crosses back to the client.
- Reach another account. The connector answers only as the person who signed in.
- Seal a new key. That happens in the vault on ahel.ai.
Questions
Everything else.
What is MCP?
The Model Context Protocol is the open standard AI clients use to reach tools and data. An MCP connector is an address your AI reads tools from. Ahel's connector serves your whole account at one address.
How does the Ahel MCP connector work?
It is a remote MCP server over streamable HTTP. After the client is configured, OAuth signs you in and each authorized call reaches your Ahel workspace. Setup steps depend on the client.
Which AI clients does this work with?
The supported setup paths cover Claude, Claude Code, ChatGPT, Codex, Cursor, and VS Code with Copilot. They authenticate through browser OAuth. Claude shows cards; Cursor, Claude Code and Codex get complete text; cards in ChatGPT and VS Code are not verified yet.
Does it work with ChatGPT?
Yes, through Developer mode, on ChatGPT plans that allow custom apps: in ChatGPT's Settings, turn on Developer mode, then create an app named ahel with Ahel's address as the server URL. We have not verified Ahel's cards in ChatGPT yet.
Why does the tool list never change?
Install a hundred servers and your client still sees the same tools: installed servers are reached through tools and use, never added to the list. A tool list that grows with every install breaks prompt caching, bloats every request, and makes clients that read the list once at connect time show stale tools. The 2026-07-28 MCP revision recommends a deterministic tool list for exactly this reason.
What can my AI actually reach?
Only what your account can already reach: your workspace's apps, the catalog, and the tools of the hosted MCP servers you have enabled. Key values never cross the connector. Nothing more, and nothing from anyone else's account.
Can it spend money?
Not without you on a write: with the confirm setting on (the default), every write your AI asks for shows a card first and runs only when you press it. A read can still cost money. A read from a paid vendor spends on your key the moment your AI makes it, and Ahel's knowledge queries, web search, hosted scripts, app actions and the models in the hosted chat are paid from your workspace balance at the prices on the pricing page. Spending from the balance stops at zero, and a failed knowledge query is never charged.
Add Ahel to your AI.
One connector. Claude shows your apps as cards; Cursor, Claude Code and Codex get complete text.
