bedcoder

Drive Claude Code from your phone. bedcoder is a tiny headless daemon you run next to the official claude: it shows a pairing code, then relays the session to your phone — end-to-end encrypted, zero signup.

npm install -g bedcoder      # or: pnpm add -g bedcoder

Usage

cd your-project
bedcoder                     # new session; prints a pairing code + QR

Open https://web.bedcoder.org on your phone, enter the code, and confirm the 6-digit SAS shown on both ends. Then send prompts, approve permissions, and read replies from your phone. Back at your desk, claude --resume picks the session right back up.

| Flag | Meaning | |------|---------| | (none) | New session; connects to the default relay wss://relay.bedcoder.org. | | --relay <wss://…> | Use a different relay (e.g. your self-hosted one). | | --resume [id] | Continue this project's latest ~/.claude session (or a specific id). | | --rewind-code | Enable file checkpointing so /rewind can restore code. Off by default — it uses git "shadow repos" and needs git + a normal, writable repo; on setups without that it can stall the agent. Conversation rewind works regardless. | | --mcp-timeout <seconds> | Hard wall-clock cap on each MCP tool call (default 600 = 10 min). A hung MCP server would otherwise wait ~27 h, freezing the turn; with a cap it fails with an error and the turn recovers. Overrides the MCP_TOOL_TIMEOUT env var. | | --stall-warn <seconds> | How long a turn may go silent (tool in flight, no permission pending) before the phone gets a "may be stuck — tap Abort" notice (default 60). 0 disables the warning. | | --fake | Echo engine — pair & test the channel without Claude auth. | | --log [path] | Write a diagnostics log (off by default). Defaults to ~/.bedcoder/agent.log; pass a path to override. |

Real Claude sessions need Claude auth: run claude once to log in, or set ANTHROPIC_API_KEY. The relay is end-to-end encrypted and never sees your content or keys; you can self-host it (single Go binary) — see the project docs.

Platform support

macOS and Linux are first-class. On Windows, run the agent inside WSL2 — it's the supported path today (zero changes; claude + bedcoder share WSL's ~/.claude). Native Windows is partially supported. See docs/WINDOWS.md.

MCP tools

MCP tools run inside the official claude CLI (configure them in ~/.claude.json or a project .mcp.json as usual). Because they call out to external processes, they can hang — and the CLI's default per-call timeout is effectively infinite, which would freeze the turn forever (the phone just keeps showing the timer). Two layers guard against that:

• Hard timeout — each MCP tool call is capped at --mcp-timeout (default 10 min). A stuck call fails with an error and the turn recovers on its own. It's a blunt wall-clock cap, so the default is high to avoid cutting off a legit slow tool; lower it if your tools are quick.
• Stall watchdog — if a turn goes --stall-warn seconds (default 60) with no activity while a tool is in flight, the phone gets a "may be stuck — tap Abort" notice naming the tool, so you can cancel by hand long before the hard cap. Set --stall-warn 0 to turn it off.

There are testing fixtures (a tiny stdio MCP server with a hang_forever tool) under test-fixtures/.

Diagnostics log

Logging is off by default. Run with --log to record a JSON-lines trace of the session (engine lifecycle, status changes, messages in/out) — useful when reporting a bug:

bedcoder --log                       # → ~/.bedcoder/agent.log
bedcoder --log /tmp/bedcoder.log      # → custom path

The active path is printed in the daemon's Log: status line. To keep with the zero-knowledge design the log records only shapes and lengths — event tags, message types, and field sizes — never message text, prompts, or keys.

Equivalent env var (handy for tests / CI): BEDCODER_LOG=<path> enables it, BEDCODER_LOG=off force-disables it (and overrides --log).

License: GPL-3.0-or-later.