Use Agent Share with Claude Code

This walks through wiring Agent Share into a Claude Code repo so your phone and your agent can share things both ways. It's the hands-on companion to Channels and the agent inbox.

Before you start: run agent-share login once so your credentials are stored. Everything below assumes you're signed in.

1. Choose how shares reach the agent

• Check on demand (recommended). Works on any Claude Code, no flags. The agent looks at your phone shares when you ask it to.
• Channels (research preview). Shares pop into the running session on their own. Needs Claude Code v2.1.80+ and an Anthropic login. The feature is still evolving.

If you're not sure, start with check-on-demand. The channels page explains the difference in plain terms.

2. Path A — check on demand (recommended)

Pick one of two ways to wire the repo. Both write a ./.mcp.json and drop a SKILL.md that tells the agent when to look at your phone.

Your personal phone inbox — read what you send from your phone and buzz it back:

agent-share init

A dedicated inbox for this agent — a named place to share to, separate from your personal inbox:

agent-share install

Either file is safe to commit — there's no token in it; auth comes from your stored login. (Add --force to overwrite an existing agent-share entry.)

Using it in a session

Just mention your phone, and the agent checks the inbox, acts on what you sent, and can reply back to your phone:

check my phone

The agent won't be interrupted by an incoming share — it picks things up when you point it there. (If you want shares to arrive on their own, use channels below.)

3. Path B — channels (research preview)

Research preview. Needs Claude Code v2.1.80+ and an Anthropic login (not Bedrock, Vertex, or Foundry). The feature may change.

With channels, anything you share from your phone — plus tasks handed off from other agents — shows up in the running session automatically.

Set it up:

agent-share register "my session"

This enrolls and names an inbox, wires the repo for channel mode pinned to it, and prints the launch line. Handy flags:

• --role "<what this session is doing>" — so you can tell agents apart when sharing from your phone.
• --dir <name> — set it up in a separate working folder.

The register file contains a token for that inbox — don't commit it. (The init / install files have no secret and are safe to commit.)

Launch Claude Code against the channel:

claude --dangerously-load-development-channels server:agent-share --dangerously-skip-permissions

Now share from your phone while the session runs — it appears in the conversation without you asking.

Already have an inbox wired and just want to switch it to channel mode?

agent-share install channel

4. Manage your inboxes

List them, with online/offline status:

agent-share agents          # active inboxes
agent-share agents --all    # include hidden ones

Hide an inbox from the phone's share picker without losing it (and its history), or bring it back:

agent-share disable "my session"
agent-share enable "my session"

You can refer to an inbox by its name or its id.

5. Point at a local dev server

If you're running the server locally under Aspire instead of production:

agent-share install aspire        # uses the local "share" resource

This wires the repo to your local server and keeps its login separate from your production one, so the two never clash. Set USE_ASPIRE=1 to make every agent-share command use the local setup.

Troubleshooting

• "It made me log in again." If your saved login has expired, agent-share signs you back in automatically in a terminal; in a non-interactive context it prints what to run.
• Is ./.mcp.json safe to commit? Yes for init and install; no for register / channel mode (those embed a token).

See also

• Channels and the agent inbox — the concepts.
• CLI reference — every command and flag.
• MCP tools reference — the tools an agent calls.