跪拜 Guibai
← All articles
Frontend · DeepSeek · Claude

Agents Anywhere Puts Claude Code, Codex, and DSH on Your Phone: A Mac mini Test

By jerrywus ·
Read original on juejin.cn ↗ Google Translate ↗ Alt translation

A long agent run no longer has to stall on a "should I execute?" prompt until someone walks back to the desk; the same session can be answered from a phone. The trade-off is operational: session data may transit a hosted server, and anyone with sensitive projects needs to plan around self-hosting or accept that boundary.

Summary

Agents Anywhere keeps the heavy machinery on a work device. Claude Code, Codex, or DeepSeek Harness runs on the Mac or server, while a Connector service exposes the session, files, and terminal to desktop, mobile, and web clients. The phone is a control surface, not a second development environment; closing it doesn't stop the agent, and reopening the session resumes the same timeline. The work device must stay online, and a sleeping or crashed machine still takes everything down.

A real run on a Mac mini showed the value. Claude Code analyzed the disk and found about 27GB of reclaimable space, with 17.6GB recommended as the priority, while correctly advising against deleting fresh Rust/Tauri build directories. The results, including tables and paths, appeared on the phone, and the "should I execute?" prompt could be answered from there. A caveat: AI can list candidates, but regenerable caches cost time to rebuild, so a human should still review paths and impact before approving.

The workbench also consolidates session management across multiple agents and machines, which matters once a developer juggles Codex, Claude Code, and DSH. It does not merge their contexts or make agents collaborate; it only gives a unified place to find a device, a project, and a session. Costs follow the original agent accounts, and session data may pass through or be stored on the server, so sensitive work pushes toward the MIT-licensed Docker self-hosting option.

Takeaways
— Agents Anywhere supports OpenAI Codex, Claude Code, and DeepSeek Harness (DSH) through one workbench.
— In a disk-analysis test on a Mac mini, Claude Code reported roughly 27GB of cleanup candidates and recommended prioritizing 17.6GB, while holding back Rust/Tauri build directories.
— The phone or browser only connects to the agent; the actual execution, files, and terminal stay on the work device.
— A Connector service runs on the work device; the desktop client includes it, and headless Linux machines use a standalone Connector CLI.
— The work device must stay powered on and networked, with both Connector and Runtime running, for mobile access to work.
— Sessions are consistent across devices: opening the same session on a phone shows the same timeline, files, and approval prompts.
— The workbench includes file browsing, preview, attachment upload/download, a remote shell, and an interactive terminal on the work device.
— Runtime capabilities such as model selection, approval, interruption, and resume can differ between agents; the client display is authoritative.
— Model costs are billed through the existing agent accounts and are not replaced or covered by Agents Anywhere.
— Session content, timelines, code snippets, and attachments may pass through or be stored on the server; Docker self-hosting is available under the MIT license.
— The recommended first test is a read-only task: start an analysis on the computer, find the same session on the phone, ask a follow-up question, then verify on the desktop.
— For Claude Code-only users, Claude's own Remote Control may be enough; the multi-agent workbench earns its keep when several agents or several machines are in play.
Conclusions

An awkward "should I execute?" wait is what remote access removes, but session organization matters just as much: with several agents and machines, finding which window belongs to which task costs more time than the remote keystroke itself.

Multi-agent support unifies access only. It does not merge Claude's context with Codex's, and the human still decides which agent fits a task.

Disk-cleanup suggestions like the 27GB figure only become useful with a human cost check; a cache can be regenerated, but regeneration time is a real price, so "reclaimable" and "free" are different categories.

The privacy section draws a line that's easy to blur: execution happens locally, but session data can transit a server, so self-hosting is a data-boundary decision rather than a deployment preference.

Concepts & terms
Work device
The computer or server that actually runs the coding agent and holds the project files and execution environment. Phones and browsers only connect to it, not replace it.
Runtime
In Agents Anywhere's terminology, the agent implementation being managed, such as Claude Code, OpenAI Codex, or DeepSeek Harness, plus the components that run or connect it.
Connector
A service on the work device that connects the agent, files, and terminal to the workbench. The desktop client includes it; headless Linux machines use a standalone Connector CLI.
Session
The conversation and execution record for one task. Reopening a session on another device shows the same timeline and lets you continue from where the agent stopped.
Source: juejin.cn ↗ Google Translate ↗ Backup ↗