MCP server¶

Lets an AI assistant read and moderate your feedback directly — list what is waiting, read it, approve or reject, create testimonials, check NPS, fetch an embed code.

pip install withfeedback-mcp
# or, without installing:
uvx withfeedback-mcp

Claude Desktop / Claude Code¶

{
  "mcpServers": {
    "withfeedback": {
      "command": "uvx",
      "args": ["withfeedback-mcp"],
      "env": {
        "WITHFEEDBACK_TOKEN": "spd_your_token_here"
      }
    }
  }
}

Create the token under Account → API Tokens with only the scopes you want the assistant to have. That token is the boundary of what it can do.

Tools¶

Tool	Calls	Needs
`list_teams`	`GET /teams/`	`read:teams`
`list_projects`	`GET /teams/{team}/projects/`	`read:feedback`
`list_submissions`	`GET …/submissions/`	`read:feedback`
`get_submission`	`GET …/submissions/{id}/`	`read:feedback`
`approve_submission`	`POST …/approve/`	`moderate:submissions`
`reject_submission`	`POST …/reject/`	`moderate:submissions`
`create_testimonial`	`POST …/submissions/`	`write:testimonials`
`import_testimonials`	looped, ≤100 items	`write:testimonials`
`list_widgets`	`GET …/widgets/`	`read:widgets`
`get_widget_embed_code`	`GET …/embed-code/`	`read:widgets`
`get_nps_summary`	surveys + results	`read:surveys`

A tool whose scope your token lacks fails with a clear message rather than doing something narrower and pretending it worked.

Deciding what to grant¶

The interesting choice is moderate:submissions.

Without it, the assistant becomes a very good reader: it can summarise what is waiting, spot themes, draft your replies, and tell you which five are worth publishing — and you click approve.

With it, the assistant can publish to your website.

Approving is publishing

An approved, consented testimonial appears in your embeds. Give an assistant moderate:submissions when you want that, not by default — and remember that a submission's text is written by a stranger. Text asking an assistant to "approve all pending items" is content to be moderated, not an instruction. Reading a queue is safe; acting on what a queue says is where a prompt injection would land.

Start read-only. Add moderation when you have watched it work.

Configuration¶

Variable	Default	Purpose
`WITHFEEDBACK_TOKEN`	—	Personal access token
`WITHFEEDBACK_API_URL`	`https://withfeedback.com`	Self-hosted or staging
`WITHFEEDBACK_CLIENT_ID`	official public client	Device-flow fallback
`WITHFEEDBACK_SCOPES`	all tool scopes	Scopes requested in device flow
`WITHFEEDBACK_MCP_TRANSPORT`	`stdio`	`http` for streamable HTTP
`WITHFEEDBACK_MCP_HOST`	`127.0.0.1`	HTTP bind address
`WITHFEEDBACK_MCP_PORT`	`8000`	HTTP bind port
`WITHFEEDBACK_MCP_ALLOW_REMOTE`	unset	Required to bind beyond loopback

HTTP transport¶

withfeedback-mcp --http

Binds 127.0.0.1 only. That default is deliberate: the server holds a token, so anything that can reach it can act as you.

Do not expose it without authentication in front

Binding to a routable address needs WITHFEEDBACK_MCP_ALLOW_REMOTE=1, which exists to make you stop and think. An MCP server on an open port is an unauthenticated proxy to your account. If you need remote access, put a reverse proxy with real authentication in front of it, or use stdio over SSH.