What the Recharm connector can do, and what it can’t

“Write” here means publishing a page to an unlisted Recharm share URL. No tool modifies or deletes existing data. Denying the two write tools at the client makes the connector read-only.

• Allow or block the connector centrally

  Claude Team and Enterprise admins decide which connectors members can add. Claude Code accepts a managed settings file that allowlists MCP servers by name, so the connector can be rolled out to a group or kept off entirely.

• Block individual tools

  Every tool has a stable name. A client-side deny rule on the two publish tools turns the connector into a strictly read-only integration without any change on Recharm’s side.

• Every call can be approved by a human

  By default Claude asks before each tool call. Publishing a brief is a distinct, named action that the user sees and confirms. Nothing is published silently.

• Per-user OAuth grants you can revoke

  Each person authorises with their own Recharm login. Removing a user from the Recharm account ends their access. No shared API keys exist to leak or rotate.

• Scope follows existing Recharm permissions

  The connector cannot widen anyone’s access. Brand-level permissions set in the Recharm app are the ceiling for what the AI can see.

• One egress domain

  All connector traffic goes to mcp.recharm.com over HTTPS. Clip previews and share pages are served from Recharm’s storage. Nothing else needs to be on an allowlist.

{
  "enabledMcpjsonServers": ["Recharm"],
  "permissions": {
    "deny": [
      "mcp__Recharm__save_brief",
      "mcp__Recharm__save_html_file"
    ]
  }
}

Is this an AI tool that can take actions on our systems?

No. The AI runs inside your approved Claude client. The connector only gives it a way to search a Recharm clip library and publish a brief back to Recharm. It has no access to your network, devices, identity provider or other SaaS.

What is the worst case if a user’s Claude session is compromised or manipulated?

The attacker could read the ad footage that user can already see in Recharm, and publish a brief or report to an unlisted Recharm URL. They could not delete anything, change settings, reach other brands or exfiltrate data to a third-party destination. Revoking the user’s access in Recharm closes the window.

How sensitive is the data?

The library contains advertising footage: creator videos, product b-roll and finished ads made to be shown publicly on Meta, TikTok and YouTube, and indexed in the Meta Ad Library and TikTok Creative Center. It holds no customer data, financial data, credentials or source code. The only personal data the connector returns is the connecting user’s own email.

Who can open a published brief?

Anyone with the link. Links are unlisted and unguessable, and they contain what the user chose to put in them: scene descriptions and previews of ad clips. If your policy requires publishing to be off, deny the two publish tools at the client and the connector becomes read-only.

Does the connector install software on the laptop?

No. The Recharm plugin is a public, MIT-licensed repository of markdown skill files and a three-line config that points at the hosted server. There is no binary, no background process and no local credential store beyond the OAuth token your Claude client already manages.

Does Recharm see our Claude conversations?

Only what the AI sends as tool inputs: search phrases, brand names and the contents of a brief being published. Chat history, attached files and other tool results stay in your Claude client. How Anthropic handles that data is governed by your Claude plan.

How do we audit what was done?

Every tool call is visible in the user’s Claude session. Published briefs are listed in the Recharm app under the brand they belong to. For a log of connector activity for your account, contact us at the address below.

Can we try it before approving it?

Yes. Install the plugin against a trial Recharm account with placeholder footage. After the OAuth step your Claude client lists all nine tools with their descriptions and input schemas, so you can inspect exactly what is exposed.