daemons-mcp

Connect Claude Code to Claude voice mode. Give the Claude mobile app what it's missing: everything you've been doing in Claude Code.

My context and capabilities were all in Claude Code on my server, but the Claude app couldn't see any of it, so I couldn't think out loud with voice mode about what I was actually building. This connects the two.

It runs next to Claude Code and gives the Claude app (phone, web, desktop) a connector that can:

• Find your context. See your recent sessions, read them, and search everything you and Claude said, plus your memory files and CLAUDE.md.
• Save tasks for later. When an idea turns into something to build, Claude writes a markdown file called a bounty. Open Claude Code later, run /bounties, and an agent builds it.

It can't run code, start agents or message anyone.

Quick start

You need uv and cloudflared.

uvx --from git+https://github.com/daemonpage/daemons-mcp daemons start

It indexes your sessions and prints a connector URL. In the Claude app, go to Settings > Connectors > Add custom connector and paste it. There is no sign-in step. A new tunnel address can take a minute or two to start working.

Then ask Claude: "what have I been working on?"

To pick up bounties from Claude Code, install the command once and run /bounties in any session:

uvx --from git+https://github.com/daemonpage/daemons-mcp daemons install-skill

daemons has to stay running for the connector to work, so it suits a computer that stays on.

Can I use Claude voice mode with Claude Code?

Not directly: voice mode in the Claude app can't see your Claude Code sessions. It can use connected tools, though, and Anthropic says they work the same way in voice mode as in text chat. daemons-mcp is a custom connector (a remote MCP server) that gives the app access to your sessions.

So you can ask out loud things like:

• "What was I working on yesterday in Claude Code?"
• "What did we decide about the payment retry?"
• "Save that as a task to build later."

How it compares

Tool	Best for	Difference
daemons-mcp	Thinking out loud with your Claude Code context, then saving tasks	One command, no sign-in step. Reads and saves tasks only, so it can't run anything on your computer
claude-voice	Driving Claude Code sessions by voice from your phone	Can start sessions and send them tasks. Uses OAuth and Tailscale Funnel; GPL
Remote Control	Steering a live session from the app	Official and built in, works on one session at a time rather than searching your history
VoiceMode	Talking to Claude Code at your computer	Adds speech inside Claude Code; doesn't connect to the Claude app's own voice mode
claude-historian, claudex	Searching past sessions from inside Claude Code	Run locally for Claude Code rather than as a connector for the app

If you want your phone to run code, use claude-voice. If you'd rather it couldn't, use this.

Tools

Tool	What it does
get_current_context	Recent sessions and open bounties. Claude calls this first.
search_memory	Keyword search across sessions and memory files
list_sessions	Sessions, newest first
read_session	The latest messages of one session
read_memory_file	One memory file in full
write_bounty	Save a task for an agent
list_bounties	Saved tasks by status
read_bounty	One saved task in full
save_note	Save a note that later searches can find

Bounties

Each bounty is a markdown file in ~/.daemon/mcp/bounties/, and its folder is its status: open, claimed, done or dropped. Nothing starts automatically.

daemons bounty list
daemons bounty show <id>
daemons bounty claim <id>
daemons bounty log <id> "what happened"
daemons bounty done <id> "result"
daemons bounty drop <id>

Security

Read this before connecting, because your transcripts may hold private things.

• The URL is the password. Anyone who has it can read your sessions and memory and write bounties and notes. To change it, delete ~/.daemon/mcp/secret and restart.

• What you share is sent to Claude. Whatever the connector returns goes into your Claude app conversation, under your Claude account's data settings.

• Secret removal is best effort. Common API key formats and KEY=value pairs are removed from results, but a password in plain prose will get through. Rotate any key you pasted into a chat, and use --only to limit what is shared:

  daemons start --only my-project --only other-project
  

• Tool output isn't indexed. Only what you and Claude said is indexed, along with memory files and CLAUDE.md. Subagents and tool results are left out.

• Treat bounties as untrusted. Text in an old transcript could try to steer the Claude app into writing a bad bounty, for example if you once pasted in a malicious web page. Read a bounty before an agent with broad permissions builds it. The /bounties command tells agents to stop and ask before anything risky.

• Local files stay local. The server listens on 127.0.0.1 only and reads ~/.claude without changing it. It writes only to ~/.daemon/mcp, which also holds a plain-text search index of your sessions.

A permanent URL

The quick tunnel's address changes on every restart, which means re-adding the connector. For a stable one, run daemons start --no-tunnel behind your own named Cloudflare tunnel or any reverse proxy pointing at 127.0.0.1:8765.

Development

uv venv && uv pip install -e '.[dev]'
.venv/bin/pytest

Limits

• Search matches keywords, not meaning.
• Voice mode hasn't been tested end to end by this project yet; the connector has been tested in the Claude app's text chat.
• It reads Claude Code's transcript files, which aren't a documented format and could change.

More at arthurcamara.com/daemons-mcp. Not affiliated with or endorsed by Anthropic. MIT licensed.