claude-code - 💡(How to fix) Fix [FEATURE] Relay mode: split UI from agentic loop for low-bandwidth/high-latency connections [1 participants]

Preflight Checklist

• I have searched existing requests and this feature hasn't been requested yet
• This is a single feature request (not multiple features)

Problem Statement

Claude Code's architecture requires the full conversation context — user prompts, tool call results, and LLM responses — to round-trip between the local machine and Anthropic's API on every turn. For users on slow, high-latency, or unstable connections (remote regions, mobile tethering, VPN-constrained corporate networks), this makes Claude Code effectively unusable.

Tool outputs dominate bandwidth. A single Read of a large file, a Grep across a codebase, or a Bash command producing verbose output can be tens of KB. This data is generated locally, sent to the API, and the response streams back — all over the constrained link. On a 50 Kbps connection, a single 20 KB tool result takes ~3 seconds just to upload, and this compounds across multi-turn agentic loops with dozens of tool calls.

The workaround (SSH to a remote machine) trades one problem for another. Running Claude Code over SSH to a machine with fast internet solves the bandwidth problem, but introduces keystroke latency. Claude Code is a full-screen TUI — every screen redraw (streaming tokens, tool output rendering, status bar updates, spinner animation) must travel back as terminal escape sequences. Even mosh can't help: its local echo prediction only works for simple shell prompts, not full-screen alternate-screen-buffer applications. Mosh falls back to pure server-side rendering for TUI apps, and its state-synchronization protocol is still gated by RTT for acknowledgments (~2 updates/sec on 500ms RTT).

Proposed Solution

A client-server split ("relay mode") where:

A thin local client runs on the user's machine, handling only:

• User input (responsive, zero-latency typing)
• Displaying streamed LLM text output
• Permission prompts (approve/deny tool use)

A relay server runs on a machine with fast internet (cloud VM, workstation, etc.), handling:

• The full agentic loop (tool calls execute locally on the relay machine)
• API communication with Anthropic (fast, low-latency)
• Workspace file access (codebase lives on the relay machine)

The wire between them carries only:

• Upstream: user prompts (small, text only — typically < 1 KB)
• Downstream: LLM text output and permission requests (semantic content only — no escape sequences, no tool result payloads, no framebuffer diffs)

UX could look like:

claude --relay-server --port 9847          # on the remote machine
claude --relay-connect host:port           # on the local machine

Transport could run over an SSH tunnel for auth/encryption, or use its own TLS + token auth.

Claude Code already has building blocks for this: --input-format stream-json, --output-format stream-json, --include-partial-messages, and permission mode flags.

Alternative Solutions

Priority

High - Significant impact on productivity

Feature Category

Performance and speed

Use Case Example

Developer on a 50 Kbps / 500ms RTT connection (rural area, travel, VPN-constrained corporate network) needs to use Claude Code against a codebase on a remote server.

Current experience: Direct Claude Code is unusable — each agentic turn uploads growing context over the slow link. SSH to the server makes the TUI sluggish (every streaming token, every tool output renders as terminal escape sequences traveling back). Even mosh doesn't help because it can't predict full-screen TUI redraws.

With relay mode: User runs claude --relay-server on the remote machine, sets up an SSH tunnel, and runs claude --relay-connect localhost:9847 locally. Typing is instant. Tool calls (file reads, greps, bash commands) execute on the remote machine and never cross the slow link. Only the LLM's text responses stream back — roughly 100x less bandwidth than the current architecture.

Additional Context

Why "just get better internet" isn't an answer:

• Many developers work from locations where fast connections aren't available — rural areas, developing regions, travel, satellite internet
• Enterprise VPNs often impose bandwidth caps and add 100-200ms latency that users can't bypass
• Agentic workloads amplify the problem: 5-30 tool calls per task, each adding a round trip, making marginal connections completely impractical

Related issues:

• #53719 — CLI-as-remote-control-client (closest architectural neighbor: proposes attaching a local CLI to a remote session, but motivated by terminal preference and assumes a fast link between client and server; doesn't address bandwidth/latency on the wire)
• #25570 — iOS thin client via claude --serve (similar split UI/backend pattern, but motivated by mobile access; tool execution stays on the user's local machine)
• #49790 — Persistent remote sessions surviving disconnect (complementary)
• #22408 — High-latency SSH rendering fix (addressed flickering, not the fundamental bandwidth problem)
• #49657 — Multi-device session sync (overlapping motivation)