Feature Request: Claude Code Mobile Companion — Remote Session Control & Voice Input
Preflight Checklist
- [x] I have searched existing requests and this feature hasn't been requested yet
- [x] This is a single feature request (not multiple features)
Problem Statement
Claude Code's agentic workflows are powerful but not fully autonomous. Agents periodically pause for input: permission to proceed, a choice between approaches, or a small clarification. Today, if you step away from your laptop — to the gym, on a hike, running errands — your agents sit idle until you return. The bottleneck isn't compute or intelligence. It's physical proximity to a keyboard.
This creates a frustrating gap: Claude Code can work for hours on your behalf, but a 10-second approval can block it for hours if you're not at your desk.
User Story:
I'm running Claude Code on my laptop with multiple agents refactoring a large codebase. I head to the gym. Fifteen minutes in, I get a notification on my phone: "Agent needs approval — proposing to restructure the database schema. Review changes?" I glance at the diff summary, tap "Approve with note," and dictate: "Looks good, but keep the existing column names for backward compatibility." The agent continues. I finish my workout. By the time I'm home, the refactor is done.
Proposed Solution
A mobile interface (tab within the Claude mobile app, or a standalone lightweight companion) that connects to active Claude Code sessions running on your machine. Two core capabilities:
1. Remote Session Monitoring & Approval
- Live session feed: See what your agents are working on, their current status, and any pending requests for input.
- Quick actions: Approve, reject, or redirect with a tap. Pre-built response options like "Yes, proceed," "Try a different approach," or "Skip this for now."
- Diff review: View proposed code changes in a mobile-friendly format. Approve or request modifications.
- Push notifications: Get alerted when an agent is blocked and needs input, so you don't have to keep checking.
2. Voice-Driven Instructions
- Voice input for directions: Dictate high-level instructions like "Focus on the authentication module next," "That approach looks good, go ahead," or "Revert the last change and try using the existing utility function instead."
- Conversational interaction: Have a back-and-forth with your agents via voice — review what they've done, ask questions about their approach, give guidance — all while walking, hiking, or commuting.
- Transcription + confirmation: Show a text preview of the transcribed instruction before sending, so nothing gets lost in translation.
Why This Matters
- Eliminates idle time: Agents keep moving even when the developer is away from their desk. A 10-second voice note replaces a 3-hour wait.
- Matches real workflow patterns: Developers think about their code constantly — in the shower, at the gym, on walks. Right now there's no way to act on those thoughts without getting back to a laptop.
- Plays to Claude Code's strength: The more autonomous the agents become, the more the human role shifts from writing code to providing direction. Direction doesn't require a keyboard — it requires a channel.
- Voice is natural for high-level guidance: The kind of input agents typically need ("yes proceed," "try X instead," "prioritize Y") is conversational, not syntactic. Voice is a perfect fit.
Technical Considerations
- Secure connection: Sessions could be linked via authenticated tunnel, QR code pairing, or through the user's Anthropic account as a bridge.
- Lightweight by design: The mobile interface doesn't need to be a full IDE. It's a control panel — status, notifications, quick input, voice. Keep it minimal and fast.
- Offline resilience: Queue instructions if connectivity drops and deliver them when reconnected.
Closing Thought
Claude Code is moving toward a world where AI agents build software while developers provide vision and judgment. That kind of oversight doesn't require sitting at a desk — it requires a communication channel. A mobile companion makes Claude Code truly asynchronous: your agents work while you live your life, and you stay in the loop with a glance and a sentence.
Alternative Solutions
_No response_
Priority
Medium - Would be very helpful
Feature Category
Other
Use Case Example
Example scenario:
- I'm running Claude Code on my laptop with 3 agents refactoring a large codebase — one handling database migrations, one updating API endpoints, and one refactoring the frontend components.
- I leave my desk to go to the gym. Five minutes into my workout, I get a push notification: "Agent 1 needs approval — proposing to restructure the database schema. Review changes?"
- I open the mobile companion on my phone, glance at the diff summary showing the proposed schema changes, and tap "Approve with note."
- I dictate via voice: "Looks good, but keep the existing column names for backward compatibility."
- The agent receives my instruction and continues working. Meanwhile, Agent 2 asks for clarification on an API naming convention. I tap a quick-reply: "Follow the existing pattern."
- I finish my workout and head home. By the time I open my laptop, all three agents have completed their work — what would have been a 3-hour idle wait was reduced to two 10-second interactions on my phone.
- This would save me hours of blocked agent time every week and let me stay productive even when away from my desk.
Additional Context
_No response_
11 Comments
Found 3 possible duplicate issues:
This issue will be automatically closed as a duplicate in 3 days.
🤖 Generated with Claude Code
This exists now — I built it.
Your "10-second approval can block it for hours" line is exactly what pushed me to build this.
https://github.com/chadbyte/claude-relay
npx claude-relayin your project directory, scan the QR code, and you get push notifications on your phone when Claude needs approval. Tap to approve from anywhere — gym, kitchen, wherever.Covers everything in your "Remote Session Monitoring & Approval" section:
Voice input isn't there yet, but it's on the roadmap (Web Speech API).
One command, no install, no cloud, no account. Free and open source. Your code never leaves your machine.
I felt the same, so I built this. C3Poh - a Telegram comms bridge for Claude Code. Your agent DMs you when it's done (or when something goes wrong), and you can DM it back in natural language to kick off new tasks (works on desktop TG or mobile).
And it's security-first. Pure stdlib Python, no inbound ports (long-polling), allowlist-based access control so only you can reach it. Community channel if you run into anything or have ideas: t.me/tinmanc3poh
Voice Input Architecture Details (from #29399, closed as duplicate)
Hey all my issue [#29399] was auto-closed as a duplicate of this one. I get the overlap on the surface, but the core of my proposal covers ground that isn't addressed here, so I want to make sure these details don't get lost. Happy to discuss any of this further.
Context
I'm a Field Application Engineer managing a fleet of embedded test devices across multiple workstations in a manufacturing environment. I use Claude Code + Remote Control daily to SSH into machines, run diagnostics, tail logs, and debug from the floor. Remote Control already shipped and it's incredible, my issue was specifically about the voice input gap that makes it impractical on mobile for anything beyond quick approvals.
What's different from this issue
This issue frames voice as one piece of a broader mobile companion concept (session monitoring, approvals, push notifications, voice). My proposal focuses specifically on server-side speech-to-text for the existing Remote Control UI, with a production-tested architecture I've been running in my own workflows. Three key areas that aren't covered here:
---
1. Progressive Chunking Pipeline (not just "transcription + confirmation")
The reliability architecture is the part that makes voice input trustworthy enough for real work:
This has been built and tested in my own production workflows for 10+ minute recordings. Without this pattern, a single connection drop on a 5-minute recording means the user loses everything and stops trusting voice input.
2. Per-User Voice Dictionary for Technical Term Accuracy
General-purpose STT models have never seen your hostnames, CLI tools, or internal project names. When I say "deploy to kv-node-7," Whisper hears "deploy to cave node seven." This is solvable:
initial_promptparameter already supports vocabulary hinting)/,.,-, unusual capitalization). Let the user tap to correct. Corrections feed back into the dictionary automatically/etc/config.yamlEven a simple "here are my custom words" text file that biases transcription would be a massive improvement for technical workflows.
3. Why Browser-Native STT Won't Work Here
I want to flag this clearly because Web Speech API has come up in community discussions around this feature. Native browser STT (Web Speech API, iOS dictation) is not viable for this use case:
Server-side transcription via a real model (Whisper variant, Anthropic's own, OS hugginface, or whatever) is the only path that works for engineering workflows where someone is dictating for 1–10 minutes hands-free.
---
The dream UX (unchanged from #29399)
Tap a mic button → put the phone in your pocket → walk around talking for a few minutes → tap stop → review the transcript with flagged terms highlighted → send. That's it. A 45-second voice message replaces 5+ minutes of thumb-typing a technical prompt.
This requires no model-level audio changes - it's purely a UI/infrastructure feature that could ship independently of any broader audio modality support.
Happy to contribute
I've built and tested this exact chunking + transcription + dictionary pattern in my own environment and know the failure modes well. If there's an opportunity to contribute code or help spec this out, I'm all in.
— @Cwilliams333
I'm going to close my issue #32779 and paste my use case here. I have a Windows computer and iPhone, so the control needs to work across mixed hardware.
Problem Statement
When I start Claude Code on my Windows desktop, all interaction is confined to
that terminal. If I step away from my desk, there is no way to see what Claude
is doing, approve permission requests, respond to questions, review plans, or
prompt further work. The session just blocks until I return.
I'm a parent of a young child and frequently away from my desk. I start Claude
Code pointed at my local codebase, step away, and come back to find it's been waiting for input for several hours. That's wasted time for both of us.
Proposed Solution
Desired workflow
Start a task in Claude Code on my Windows desktop (CLI or Desktop app)
Step away from my desk
Get a push notification on my phone when Claude needs any input —
permission approvals, plan reviews, clarifying questions, or "task complete"
From the Claude mobile app (or a web UI), read Claude's full output
(plans, results, diffs, errors)
Respond conversationally — approve a plan, ask Claude to change approach,
answer a clarifying question, or prompt entirely new follow-up work
Claude continues executing on my desktop with access to my local files
This isn't just about approving/denying permission requests — it's full
conversational control of a desktop session from another device.
This is very needed. A temporary solution I am using while I need to use voice input on the Claude IOS app is using another app (like ChatGPT IOS app) to do the dictation and paste the entire text block back to Claude app remote control session. Switching between different app is faster than using Claude app chat tab and switch back to remote control Code session.
Voice input for Claude Code Remote Control would be great.
Please add also a dedicated (sub)agent for this. I want to have a conversation with the agent that will write the prompt or approve the next step for Claude Code without polluting the context.
The agent could also summarize what Claude Code is doing or help me to get a better understanding by starting sub agents.
Thank you for Remote Control. I switched to Claude Code a few weeks ago just because of this feature. I love it.
+1 , this is a major pain point for me.
I use Remote Control daily to interact with Claude Code sessions from my phone. The experience is great except for one thing: typing detailed technical prompts on a phone keyboard is brutally slow. I find myself abbreviating, simplifying, or just waiting until I'm back at my desk, which defeats the whole purpose of Remote Control.
The Claude mobile app already has dictation and voice mode in regular chats. The fact that it's missing in Code menu / remote sessions feels like an oversight. Just enabling the same microphone button that already exists in normal chats would be a huge win.
This is genuinely the difference between Remote Control being something I use occasionally and something I rely on throughout my day.
This is exactly the problem I kept running into — agent blocks on a single confirmation and you're not at your desk. I built pikiclaw specifically for this: it bridges your local Claude Code (also Codex and Gemini CLI) to Telegram/Feishu/WeChat, so approvals, follow-ups, and file exchanges all happen from your phone while the agent runs on your machine. It also handles Codex's Human Loop — when Codex asks a question mid-task, it surfaces as an interactive prompt in your IM. Quick start:
npx pikiclaw@latestGitHub: https://github.com/xiaotonng/pikiclaw
Closing for now — inactive for too long. Please open a new issue if this is still relevant.
This issue has been automatically locked since it was closed and has not had any activity for 7 days. If you're experiencing a similar issue, please file a new issue and reference this one if it's relevant.