Summary
Add a mechanism to send a prompt to an already-running Claude Code session from an external process and receive the response — without requiring the user to type in the UI.
Use case
I'm building a WhatsApp companion bot (Hermes) that bridges my phone to my Claude Code sessions. When I'm away from my Mac, I want to send a message like "check the Saturno project for pending proposals" and have it run inside my actual active Saturno session — with its full conversation history and context — then receive the result back on WhatsApp.
Today I can run claude --print "task" --cwd /project/dir which correctly picks up CLAUDE.md and project files. But this creates a new session, losing the in-memory conversation history of the session currently open in the UI.
Proposed solution
Any of these would work:
- Named pipe / Unix socket per session:
~/.claude/sessions/<session-id>.sock — write a prompt, read the response.
--session-id <uuid> flag for --print: run a task inside an existing session by ID, not just resume its history.- Local HTTP endpoint per running session (e.g.,
localhost:PORT/prompt) — POST a message, GET the streaming response.
--continue --print that actually injects into the latest session rather than creating a fork.
Why this matters
This is the missing piece for Claude Code to work as a true background agent. Right now external tools can:
- ✅ Trigger new sessions
- ✅ Read project files
- ✅ Use
--continue to resume history
But they can't:
- ❌ Talk to the session the user has currently open
- ❌ Inject a prompt into a running agentic loop
- ❌ Get a response from a session without the user being at the keyboard
A simple IPC mechanism would unlock Claude Code as a true AI companion — reachable from mobile, from other processes, from voice assistants, from automation tools.
Current workaround
Running claude --print "task" --cwd /project/dir in the correct project directory. Works, but creates a parallel session without the active conversation context.
Summary
Add a mechanism to send a prompt to an already-running Claude Code session from an external process and receive the response — without requiring the user to type in the UI.
Use case
I'm building a WhatsApp companion bot (Hermes) that bridges my phone to my Claude Code sessions. When I'm away from my Mac, I want to send a message like "check the Saturno project for pending proposals" and have it run inside my actual active Saturno session — with its full conversation history and context — then receive the result back on WhatsApp.
Today I can run
claude --print "task" --cwd /project/dirwhich correctly picks up CLAUDE.md and project files. But this creates a new session, losing the in-memory conversation history of the session currently open in the UI.Proposed solution
Any of these would work:
~/.claude/sessions/<session-id>.sock— write a prompt, read the response.--session-id <uuid>flag for--print: run a task inside an existing session by ID, not just resume its history.localhost:PORT/prompt) — POST a message, GET the streaming response.--continue --printthat actually injects into the latest session rather than creating a fork.Why this matters
This is the missing piece for Claude Code to work as a true background agent. Right now external tools can:
--continueto resume historyBut they can't:
A simple IPC mechanism would unlock Claude Code as a true AI companion — reachable from mobile, from other processes, from voice assistants, from automation tools.
Current workaround
Running
claude --print "task" --cwd /project/dirin the correct project directory. Works, but creates a parallel session without the active conversation context.