@maisterflow/applicant-mcp

Local stdio MCP for the maisterflow creator-applicant vetting flow. It runs on your own machine and exposes four tools:

1. list_sessions — lists your local Claude Code sessions (id, project, date, message/tool-call counts, a short first-prompt snippet) so you can pick the ~5 that best show how you work. Local read only.
2. prepare_assessment(session_ids) — returns the full local transcript text of those sessions plus the evidence-package schema and the strengths (advocate) instructions. Local read only. Your assistant reads the transcripts and writes the quote-backed, positive half of the evidence package.
3. prepare_concerns(session_ids) — returns the same transcripts plus the prosecutor instructions: a separate, hostile re-read that writes a quote-backed concerns list of the applicant's weakest moments (zero is fine). Local read only. This is added to the package from step 2.
4. submit_assessment(evidence_package) — the only network call. It recomputes a deterministic v3 digest over the same sessions, stamps each quote — strengths AND concerns — against the transcript (a faithfulness flag, never a gate), redacts the quotes, stores a pending application, and returns a private browser review link.

There is no AI/model call inside the tool — the digest, quote-stamping, and redaction are plain file reads and arithmetic. The host LLM writes the evidence; the tool never authors a score and never authors the verified flag. Nothing is final until you approve it on the review page; you can also delete it there. Scoring runs server-side, after you submit — never on your machine.

Use

You get an applicant token from the Apply page of the web app. Add this to your Claude Code MCP config:

{
  "mcpServers": {
    "ai-masterclass-applicant-capture": {
      "command": "npx",
      "args": ["-y", "@maisterflow/applicant-mcp"],
      "env": {
        "APPLICANT_TOKEN": "atk_...",
        "APP_BASE_URL": "https://maisterflow.com"
      }
    }
  }
}

Then ask Claude Code to list your sessions, pick the ~5 that best show how you work, and prepare and submit your assessment. Your assistant writes a positive strengths pass and then a separate adversarial concerns pass over the same sessions. The tool returns a review link; open it in your browser to see exactly what was stored — the digest, your evidence (strengths and concerns), and two un-curated openers — then submit it for review or delete it.

Environment

| Var | Required | Default | Meaning | |-----|----------|---------|---------| | APPLICANT_TOKEN | yes | — | Your one-time applicant token (from the Apply page). | | APP_BASE_URL | yes | — | The web app base URL. | | TRANSCRIPTS_ROOT | no | ~/.claude/projects | Absolute path to your Claude Code history, if it is not in the default location. |

Scope / consent

You choose ~5 sessions to feature. Your assistant reads their full local text and writes a quote-backed summary, but only the redacted evidence and the deterministic digest leave your machine — the full session text never does. The browser review page shows exactly which sessions, totals, signals, the whole-corpus anchor, the quote-backed evidence — strengths and the adversarial concerns pass, each with a verified/unverified badge per quote — and two un-curated openers, before you decide whether to submit. Nothing is hidden from you: you see the concerns your assistant raised and can delete the application instead of submitting. Pending applications expire after 7 days if untouched.