Attribura

Connect an agent

Beta

Attribura speaks MCP over a hosted server — there is nothing to clone, build or keep running. Connect once and an agent can read your brand, see which posts made money, and push finished content back for you to publish.

One command

Mint a token on the MCP page and it prints this line with your token already in it. The raw token is shown exactly once, so copy the line before you leave the page.

bash
claude mcp add --transport http attribura \
  https://api.attribura.com/mcp \
  -H "Authorization: Bearer atb_YOUR_TOKEN"

The transport is Streamable HTTP, so any MCP client can use the same URL and header — Claude Code is simply the one with a one-line installer.

Full access vs read only

A token carries one of two scopes, and the difference is a single rule rather than a permission list: a read token may only make GET, HEAD and OPTIONS requests. Anything else is refused with a 403before it reaches the handler. That one rule is where every “this tool needs full access” below comes from — a write tool fails because of the method it uses underneath, not because someone maintains a second list of which tools are safe.

Full access

All 56 hosted tools. The agent can save your brand, push creatives and edit the library. The token for content work.

Read only

The 35tools that only read. Safe to drop in a script or hand to an agent you don't fully trust — it cannot change anything, on any endpoint.

A read token still connects to the same server and still LISTS all 56 tools; the write ones answer 403 when called. The tool catalog marks which is which.

Check it worked

  1. Run /mcp inside Claude Code. attribura should be listed as connected, with 56 tools in 8 domains.
  2. Ask it something only your account can answer — “use the attribura MCP and call get_overview”. You should get your own revenue back, not an auth error.
  3. Come back to the MCP page. The token you just used should have stopped saying Never used — that stamp is written on every authenticated call, so it is the one check that cannot be fooled by a client that merely looks connected.

When it doesn't

  • connection failed — usually the header was mangled by the shell. Keep the quotes around Bearer ….
  • 401 on every call — the token was revoked. Flipping the scope toggle on the MCP page mints a replacement and revokes the token it just replaced, so a command line you pasted a minute earlier can be dead. Run the new line.
  • 403 on writes only — not a bug. You minted a read-only token and it is doing its job.
  • A tool the agent cannot find — check the catalog: some tools are written but not released, and two more are local-disk only, so they exist on the stdio server and never on the hosted one.