AI Agent Access (MCP)

FlareBuilder exposes a Model Context Protocol (MCP) server so AI agents — Claude, or any MCP-compatible client — can read and draft content in your organization directly, using your template schemas as their guide.

Caution

Agent tokens can only ever create drafts. Nothing an agent creates is published automatically — it goes through the same human review/publish workflow as content created by a person. Agents cannot publish, edit existing content, delete content, or manage anything tenant-wide.

Enabling MCP Access

Two switches must both be on before the /mcp endpoint accepts requests for your organization:

1. A tenant admin enables it tenant-wide: Settings → AI Agent Access → Enable MCP Server.
2. Each individual user who wants to connect an agent creates their own personal agent token from their Agent Tokens page (editors and above only — contributors can’t issue tokens).

Note

MCP access is gated behind the mcp_agent plan feature. It’s included on all current plans, but if your organization is on an older plan and the toggle is unavailable, you’ll need to upgrade.

Agent Tokens

A token belongs to a single user and is shown once, at creation — format fb_agt_.... Store it like a password; FlareBuilder only ever stores its hash.

• Up to 5 active tokens per user
• Every token is scoped to a subset of the creating user’s own privileges — today the only grantable scope is create draft content, regardless of how much the issuing user can otherwise do
• Revoking a token (or demoting its owner below editor) invalidates it immediately

Tip

Because a token can never be scoped beyond draft creation, giving an agent a token is safe even for your most trusted editors — the blast radius of a leaked or misbehaving token is “some unpublished drafts appear,” never “content changed or disappeared.”

Connecting an Agent

The server is a single stateless HTTP endpoint per organization:

POST https://your-org.flarebuilder.com/mcp
Authorization: Bearer fb_agt_...

Claude Code

claude mcp add --transport http flarebuilder \
  https://your-org.flarebuilder.com/mcp \
  --header "Authorization: Bearer fb_agt_..."

Claude Desktop / mcp-remote

Claude Desktop doesn’t yet speak streamable HTTP directly — bridge it with mcp-remote in your MCP config:

{
  "mcpServers": {
    "flarebuilder": {
      "command": "npx",
      "args": [
        "mcp-remote",
        "https://your-org.flarebuilder.com/mcp",
        "--header",
        "Authorization: Bearer fb_agt_..."
      ]
    }
  }
}

Generic HTTP client

Any MCP client that supports the streamable HTTP transport can connect directly — point it at the URL above with the Authorization header set. No session negotiation is required; each request is authenticated independently.

Available Tools

Tool	Purpose	Requires draft-create scope?
list_templates	List enabled templates for the organization — call this first	No
get_template_schema	Get a template’s full field schema (sections, fields, types, validation)	No
list_tags	List existing tags, to reuse instead of inventing new ones	No
list_channels	List feed channels, for create_content_draft’s channel_id	No
create_content_draft	Create a new content item as a draft	Yes
upload_media	Upload a base64-encoded image/file, returns a URL for use in a draft’s fields	Yes
list_content	List content items, filterable by template, tags, or status	No
get_content	Get a single content item by ID, including all field values	No
search_content	Full-text relevance search over content (title and body)	No

Note

Read access (list_content, get_content, search_content, list_tags, list_channels) reflects the token owner’s normal visibility in the workspace — for an editor or admin, that’s typically the whole organization’s content, not just their own. Write access is always locked to draft creation, no matter how much the owner could otherwise do. search_content searches every status (draft, published, scheduled, expired) by default, so an agent can check for duplicates before calling create_content_draft. list_channels never includes a channel’s access token, even for admin-owned agent tokens — it’s a bearer secret an agent has no legitimate use for.

Typical Flow

1. list_templates — discover valid template_id values.
2. get_template_schema — read the field schema for the template you’re targeting, so field values match the expected sections/types.
3. list_tags / list_channels (optional) — reuse existing tag names and pick a channel_id instead of guessing.
4. upload_media (optional) — upload any images first, and use the returned URLs in the next step.
5. create_content_draft — create the item with template_id, title, and metadata shaped as { section_id: { field_id: value } }, matching the schema from step 2.

// create_content_draft input
{
  "template_id": "tmpl_abc123",
  "title": "Q3 Product Update",
  "metadata": {
    "content": {
      "description": "A summary of what shipped this quarter.",
      "body": "<p>...</p>"
    }
  },
  "tags": ["news", "product"],
  "channel_id": null
}

The response never includes a date_published — drafts are always unpublished until a human reviews and publishes them from the FlareBuilder editor.

Errors

Error	Meaning
Missing bearer token (401)	No Authorization: Bearer ... header was sent
Invalid or unauthorized token (401)	Token is malformed, revoked, expired, or belongs to a different organization
MCP agent access disabled (403)	The tenant admin has not enabled Settings → AI Agent Access
AI agent access (MCP) is not included in your current plan (403)	Your plan doesn’t include the mcp_agent feature — upgrade to enable it
This agent token is not scoped for {privilege}	The tool called requires a privilege this token wasn’t granted

Tip

See Templates for the full field-type and schema reference get_template_schema returns, and Webhooks if you want to react to content once a human publishes what an agent drafted.