Copilot Digest Assistant

For QwenPaw documentation, please refer to: https://qwenpaw.agentscope.io/docs/intro

Blog post: Beyond the Chat Box: What a Personal AI Assistant Actually Is (中文)

Demo

Click to watch on YouTube (~1.5× speed, subtitled).

---

Connect your QwenPaw Copilot Digest assistant to claude.ai (web) or the Claude mobile app in about 10 minutes.

After setup, you can:

• Get ranked briefings: "what's new today?", "catch me up on this week"
• Browse your reading list: "what's in my reading list?"
• Read full articles: "tell me about the Nvidia paper"
• Save URLs: "save https://example.com to my reading list"
• Save files: "ingest ~/papers/attention.pdf"
• Discuss articles and capture notes: "discuss #3", "save these notes"
• Export briefings: "export today's briefing"
• Manage interests: "add AI policy to my topics", "what sources am I tracking?"
• On mobile, ask questions about your reading list via Claude

---

Setup overview

• Agent id: <your-agent-id> (Copilot Digest) — find via qwenpaw agents list
• Workspace: ~/.qwenpaw/workspaces/copilot-digest/
• Token file: ~/.qwenpaw/mcp_token
• Ports: QwenPaw backend 8088, MCP server 8089

Sanity-check the agent and grab its id:

qwenpaw agents list

Also needed: cloudflared (brew install cloudflared) and a claude.ai Pro / Team / Enterprise account (custom connectors are paid-tier).

---

Step 1 — Start QwenPaw

qwenpaw app

Leave this running. Expected: QwenPaw logs "Uvicorn running on …:8088".

Step 2 — Start the MCP server

In a new terminal:

qwenpaw-mcp serve --no-auth --agent-id <your-agent-id> \
  --workspace ~/.qwenpaw/workspaces/copilot-digest

The --workspace flag enables direct workspace tools that read/write your knowledge base files directly without an assistant round-trip. Without it, only assistant-proxied tools are available.

Expected output on stderr:

MCP bearer token: qp_mcp_abc123...
Stored at: /Users/you/.qwenpaw/mcp_token
qwenpaw-mcp listening on http://127.0.0.1:8089 ...

Copy the token. You'll paste it into claude.ai in step 4. (It's also saved to ~/.qwenpaw/mcp_token for later — cat that file anytime.)

Step 3 — Open a public tunnel

In a third terminal:

cloudflared tunnel --url http://127.0.0.1:8089

Expected output includes a line like:

Your quick tunnel has been created!
https://lovely-possum-42.trycloudflare.com/mcp

Copy that HTTPS URL.

⚠️ Keep all three terminals running while you use the connector.

Step 4 — Register in Claude

1. Go to claude.ai → Settings → Profile → Connectors → Add custom connector.
2. URL: paste the https://…trycloudflare.com URL from step 3.
3. Authentication: choose Bearer token, paste the qp_mcp_… token from step 2.
4. Save. Open a new chat and enable the connector.

Step 5 — Try it

In a claude.ai chat with the connector enabled, say:

Save this URL to my reading list: https://example.com

Claude should invoke send_message. You'll see a reply like "Saved." and a new entry should appear under ~/.qwenpaw/workspaces/<…>/articles/ on your machine.

On Claude mobile:

1. Open the app → Settings → Connectors — your new connector should already be there (it syncs from claude.ai).
2. Enable it.
3. Ask "what's new on my reading list?" to get started.

---

Tools

Direct workspace tools (fast, no assistant round-trip)

These require --workspace to be set. They read/write workspace files directly and respond in milliseconds.

Browse & read

Tool	What to say to Claude
list_reading_list	"what's in my reading list?", "show unread articles", "what did I save this week?"
get_article	"tell me about the Nvidia paper", "show me article abc123" (returns the curated _script.md summary when available, otherwise the raw article)
get_stats	"how many articles do I have?", "knowledge base stats"

Briefings

Tool	What to say to Claude
get_briefing	"what's new today?", "this week's briefing", "weekly briefing", "catch me up", "unread items on fintech"
export_briefing	"export today's briefing", "compile my notes and articles"

Status tracking

Tool	What to say to Claude
mark_read	"mark that as read", "I've read article abc123"
mark_unread	"mark that as unread again"
mark_discussed	"we discussed this" (also marks as read)

Work outputs

Tool	What to say to Claude
save_work_output	"save my notes", "save my discussion notes", "record the takeaways", "write down the action items"

Configuration

Tool	What to say to Claude
get_config	"what sources am I tracking?", "show my topics", "what's my fetch schedule?"
update_config	"add AI policy to my topics", "remove crypto from my interests", "add TechCrunch as a source"

Assistant-proxied tools (need LLM)

These forward to the Copilot Digest assistant and may take 10-60 seconds.

Tool	What to say to Claude
send_message	"discuss the SEC case", "draft a summary", "set up auto-fetch"
ingest_url	"save https://example.com to my reading list"
ingest_file	"ingest /Users/me/papers/attention.pdf" (path on your QwenPaw machine)

Session management

Tool	What to say to Claude
reset_session	"forget the previous conversation" / "start over"

---

How it works

Claude (claude.ai web / mobile app)
  │
  │  Streamable HTTP  (MCP 2025-03-26 transport)
  ▼
cloudflared tunnel
  │
  ▼
qwenpaw-mcp server  (port 8089, uvicorn + mcp.streamable_http_app())
  ├── server.py      — FastMCP tool definitions, handles the MCP protocol
  ├── workspace.py   — direct read/write of workspace files
  │                    (used by the fast tools: list_reading_list,
  │                     get_article, get_briefing, mark_*, get_stats,
  │                     get_config, update_config, save_work_output,
  │                     export_briefing)
  └── client.py      — HTTP + SSE proxy to the QwenPaw backend
                       (used by send_message, ingest_url, ingest_file)
          │
          │  POST /api/agents/{agent_id}/console/chat
          │  response: text/event-stream (backend's own API, not MCP)
          ▼
QwenPaw backend  (port 8088)
  │
  ▼
Filesystem  /  LLM

Two transports, don't confuse them. The Claude ↔ MCP hop uses Streamable HTTP — the current MCP transport standard, which replaced the older SSE-only transport. The MCP ↔ QwenPaw-backend hop separately uses an SSE-style response stream, but that's the backend's own console API, not the MCP protocol.

Two latencies. The fast path (workspace.py) answers browse/read/status calls in milliseconds by touching workspace files directly. The slow path (client.py) forwards conversational work to the Copilot Digest assistant, which can take 10–60 seconds because an LLM is in the loop.

---

Troubleshooting

What you see	What to try
401 Unauthorized in claude.ai	Token mismatch. Re-copy ~/.qwenpaw/mcp_token into the connector settings.
Cannot reach QwenPaw at …	qwenpaw app isn't running. Start it.
Agent '<your-agent-id>' not found	Agent was deleted or recreated with a new id. Run qwenpaw agents list and update --agent-id.
Console channel not registered on agent …	The agent was created without the console channel. Recreate it with qwenpaw agents create ….
QwenPaw did not finish within 180s	Long PDF ingest. Raise --timeout 300 when starting the MCP server.
Connector shows up but tool calls hang	Check Terminal 2 (MCP server) for errors. Restart the tunnel — trycloudflare.com URLs sometimes drift.

---

Rotating the bearer token

If you think the token leaked:

rm ~/.qwenpaw/mcp_token
qwenpaw-mcp serve --agent-id <your-agent-id> --print-token

A new token is generated on first run. Update the claude.ai connector with the new value.

---

Stop everything

• Ctrl-C in each of the three terminals.
• The claude.ai connector stays registered; next time just start the three processes again and the existing token still works.

---

Want more detail?

• Copilot Digest skill behavior: src/qwenpaw/agents/skills/copilot_digest/SKILL.md
• HTTP contract we proxy: src/qwenpaw/app/routers/console.py