MCP server

Avtrz ships a remote MCP server, so Claude, ChatGPT, and any MCP client can resolve real business profile photos and enriched profile data inline — no API key, no glue code. Add one connector URL, authorize once, and pick the workspace it bills to.

https://avtrz.dev/mcp

The Avtrz MCP server speaks the Model Context Protocol over streamable HTTP. https://avtrz.dev/mcp is the connector URL you paste into your client — not an endpoint you POST to yourself; the MCP transport handles the calls. It exposes the same avatar engine behind GET /v1/avatar as a set of read-only tools an LLM can call mid-conversation. Usage meters the same workspace quotas — there is no separate MCP billing.

What it is

A connector your AI client adds once. When a model needs the face or the facts behind a LinkedIn URL, it calls an Avtrz tool and gets the answer back inline:

• get_profile_avatar — the real profile photo as an inline image (or deterministic initials art while enrichment runs), plus a ready-to-paste imageUrl.
• get_profile — structured person data (name, headline, summary, industry, location, picture URL, open-to-work) plus a readable Markdown block.
• get_publishable_key — your workspace's publishable avatar key and a URL template, so the model can hand you durable <img> URLs that render anywhere.

All three tools are read-only. The connection can't write data, change billing, or message your team.

Working image URLs

The MCP server doesn't just return image bytes — it can hand the model a durable Avtrz image URL to drop into a CRM, doc, email, or static site, where it renders immediately:

https://avtrz.dev/v1/avatar?key=pk_…&linkedin_url=https://www.linkedin.com/in/janedoe&size=128

These URLs are powered by your workspace's publishable key, provisioned automatically the first time the connector is used. get_profile_avatar returns one as imageUrl alongside the inline image, and get_publishable_key returns the key plus a URL template so the model can build URLs for a whole list of people at once. The key is unrestricted by design (no domain firewall) so the images load from any context — see MCP tools for the full reference and caveats.

How you connect

• Add the connector URL https://avtrz.dev/mcp in your client (Claude, ChatGPT, Cursor, or any MCP client).
• The client discovers the authorization server and starts an OAuth 2.1 + PKCE handshake.
• You land on the Avtrz consent screen, sign in, and pick the workspace this connection binds to.
• Authorize. The client gets a scoped access token and the tools light up.

The consent screen auto-detects which client is connecting (Claude, ChatGPT, or a generic MCP client) from its redirect URI, and shows you the exact scopes, the read-only guarantee, and the workspace switcher. You can revoke the connection anytime from your dashboard. See Connect Claude & ChatGPT for the step-by-step.

Workspace binding, quota, and billing

The handshake binds the connection to one Avtrz organization (workspace). Every tool call is metered against that workspace, exactly like an API-key request:

• Tool calls write to the same avatarRequest analytics the HTTP API uses, so MCP traffic shows up in your dashboard alongside everything else.
• Every metered tool call counts as one avatar_requests unit — whether or not a photo is found — and a profile your workspace hasn't looked up before also counts against avatar_new_profiles.
• Hit a monthly limit and the tool returns a clear "quota exceeded for this workspace" message — no surprise charges.

Who it's for

• Sales and RM teams who live in Claude or ChatGPT and want a real face and one-line profile next to every prospect they discuss.
• Builders wiring Avtrz into an agent via Cursor or the MCP Inspector without hand-rolling an HTTP client.
• Anyone already using the GET /v1/avatar API who wants the same engine available conversationally, on the same quota.

How it relates to the API

The MCP server is a second front door to the same avatar engine — not a separate product. Reach for the API when you render avatars in your own product. Reach for the MCP server when a model should resolve them in conversation.