Paje MCP

Turn the real work you do in Claude into a credible, in-voice personal brand.

Paje listens to your actual working sessions, mines them for notable work, and turns a week of work into one brief: a single identity throughline, a posting cadence, and ranked post drafts — every one grounded in something you really did. No blank prompts, no scraping, no inventing results.

The wedge: provenance over prompts. Built from real work, it's credible, in-voice, and un-fakeable. A confidentiality filter abstracts client-sensitive material and holds risky posts for human review before anything surfaces.

This is a local-first Model Context Protocol server (stdio transport). Connect it to Claude Desktop or Claude Code.

What it does

| Tool | Purpose | | --- | --- | | capture_work | Log one notable piece of work (title, summary, decisions, tags). | | ingest_transcript | Mine a full conversation/transcript for notable work. | | weekly_brief | A week of work → throughline + cadence + ranked drafts. | | draft_posts | Ranked drafts (reach vs. risk) from captured work, in your voice. | | list_corpus | List recently captured work. | | set_voice | Set positioning, audience, and tone so drafts sound like you. | | sync_to_paje | Build the weekly brief and push it (+ public-safe work items) to your paje dashboard. Returns a private view link. |

It also exposes a paje://corpus resource and a log_session prompt.

Everything runs through a two-layer confidentiality pass: a deterministic scrubber (emails, phone numbers, API keys, env secrets, private keys, card numbers — never stored or sent to a model) plus model-side abstraction of client identities. Drafts that draw on sensitive work are marked hold_for_review.

The corpus is a single JSON file at ~/.paje/corpus.json (override with PAJE_HOME). It's the source of truth; briefs and drafts are always derived from it, never from a blank prompt.

Sync to your paje dashboard

sync_to_paje pushes your latest weekly brief — throughline, cadence, and ranked drafts — to your paje page so it shows up on your dashboard and at a private, unguessable link. The brief is built on your machine, so the confidentiality pass runs locally: items marked private are dropped entirely and sensitivity notes never leave your device. weekly_brief also auto-syncs when sync is configured, so just asking for your brief keeps the dashboard fresh.

To turn it on:

1. In paje.ai, open Dashboard → Paje. Copy the PAJE_SYNC_URL and PAJE_SYNC_TOKEN shown there.
2. Add them to the env block of the paje server in your Claude Desktop config (alongside ANTHROPIC_API_KEY), then fully quit and reopen Claude.
3. In any chat, say "sync to paje" (or just ask for your weekly brief). The tool returns the view link, and the brief appears on Dashboard → Paje.

The token is a shared secret: keep it private. Anyone with the resulting view link can read that brief (but not push to it).

Quick start (no cloning required)

npx paje-mcp setup

That's it. The CLI auto-detects your OS, finds Claude Desktop's config, prompts for your Anthropic key, and writes the config for you. Restart Claude and you're done. See paje.ai/setup for a visual walkthrough.

Install from source

If you prefer to build from source (e.g. for development):

pnpm install
pnpm --filter paje-mcp build

This produces dist/index.js (the MCP server) and dist/cli.js (the CLI).

Environment

| Var | Required | Purpose | | --- | --- | --- | | ANTHROPIC_API_KEY | recommended | Best-quality extraction & drafting. | | OPENROUTER_API_KEY | optional | Fallback model if no Anthropic key. | | PAJE_HOME | optional | Corpus dir (default ~/.paje). | | PAJE_ANTHROPIC_MODEL | optional | Override the Anthropic model. | | PAJE_OPENROUTER_MODEL | optional | Override the OpenRouter model. | | PAJE_SYNC_URL / PAJE_SYNC_TOKEN | optional | Push the weekly brief to your paje dashboard (see Sync to your paje dashboard). |

Without any model key, Paje still works using a heuristic extractor/drafter — set a key for in-voice quality.

Connect to Claude Desktop

One command (recommended)

npx paje-mcp setup

This auto-detects your OS, finds the Claude Desktop config, prompts for your Anthropic key, and writes the paje MCP server entry (preserving any other servers). Then fully quit Claude Desktop (Cmd+Q on macOS) and reopen it.

Alternatively, from inside the cloned repo:

export ANTHROPIC_API_KEY="sk-ant-..."   # optional — it'll prompt if unset
bash packages/paje-mcp/install.sh

Skip to Use it. The manual steps below do the same thing by hand.

Manual

Add Paje to your Claude Desktop config:

• macOS: ~/Library/Application Support/Claude/claude_desktop_config.json
• Windows: %APPDATA%\Claude\claude_desktop_config.json

{
  "mcpServers": {
    "paje": {
      "command": "node",
      "args": ["/absolute/path/to/paje.ai/packages/paje-mcp/dist/index.js"],
      "env": {
        "ANTHROPIC_API_KEY": "sk-ant-...",
        "PAJE_HOME": "/Users/you/.paje"
      }
    }
  }
}

Restart Claude Desktop. To make capture semi-passive, add a Project instruction:

At the end of a working session, run the log_session prompt (or call capture_work for each notable thing we did). Then I'll review my weekly brief.

Use it

Once Claude Desktop restarts, click the tools/connector icon — you should see paje. Then:

1. Paste a real work conversation and say: "Ingest this transcript into Paje."
2. Ask: "Give me my weekly brief." → throughline + cadence + ranked drafts, with anything sensitive held for review.

Connect to Claude Code (passive capture)

Claude Code can auto-capture every session on the Stop event. Add to .claude/settings.json (project or ~/.claude/settings.json for global):

{
  "hooks": {
    "Stop": [
      {
        "matcher": "",
        "hooks": [
          {
            "type": "command",
            "command": "/absolute/path/to/paje.ai/packages/paje-mcp/hooks/claude-code-stop-hook.sh"
          }
        ]
      }
    ]
  }
}

The hook receives Claude Code's session payload (incl. transcript_path), pipes it into paje capture, mines it for notable work, scrubs secrets, and appends to your corpus. It never blocks Claude — if the build is missing it exits cleanly.

Set ANTHROPIC_API_KEY (and PAJE_HOME) in your shell so the hook can see them.

CLI

Useful for manual capture and for testing without an MCP client:

# mine a transcript (or a Claude Code hook payload) from stdin
paje capture < examples/sample-session.txt

# your week: throughline + cadence + ranked drafts
paje brief --days 7

# just the drafts
paje draft --count 5 --platform linkedin

# what's captured
paje list

# push the weekly brief to your paje dashboard (if PAJE_SYNC_URL/TOKEN are set)
paje sync [--days 7]

During development, run via pnpm --filter paje-mcp capture (uses tsx), or use the built binary after pnpm --filter paje-mcp build.

How it fits the bigger picture

Paje builds five things — positioning, proof, a website you own, content, and an owned audience. This MCP is the engine under content: it watches the work and keeps the posts grounded in it. sync_to_paje is the seam that closes the loop — posts drive to the site, the site converts to the list.