SeaWorkSeaWork
简体中文
下载
2026年07月14日

How to Use Grok 4.5 / Grok Build — From Your Phone (2026)

Grok 4.5 is the default model in Grok Build. Here's how to run it for coding, what the price and benchmarks mean, and how to steer a Grok agent from your phone with SeaWork — code stays on your machine.

To use Grok 4.5 for coding from your phone, run it as the default model in Grok Build on your machine, then steer that local agent from SeaWork on iOS, Android, desktop, or the web. Install Grok Build (curl -fsSL https://x.ai/cli/install.sh | bash or let SeaWork install it for you), pick model grok-4.5, and control the session remotely while the agent — and your code — stay on your laptop. Below: what shipped, when Grok 4.5 is the right pick, three real commands, and how remote control works without moving the repo to a cloud sandbox.

That’s the short answer. The rest is the honest version — because “use Grok from your phone” is easy to confuse with a hosted chat app, and the useful pattern is the opposite: local agent, mobile control surface.

What shipped

SpaceXAI released Grok 4.5 on July 8, 2026 as its strongest model for coding, agentic work, and knowledge tasks (SpaceXAI; TechCrunch). The numbers that matter for day-to-day coding:

  • Model ID: grok-4.5 (aliases include grok-4.5-latest, grok-build-latest)
  • Pricing: $2 input / $0.50 cached input / $6 output per million tokens (model docs)
  • Speed: about 80 tokens/sec on SpaceXAI’s serving claim
  • Context: 500K tokens; text + image in → text out; function calling and structured outputs
  • Surfaces at launch: Grok Build (default model), Cursor (all plans), SpaceXAI API — not yet in the EU at launch

Grok 4.5 is a model. Grok Build is the terminal coding agent that defaults to it — plan mode, clean diffs, AGENTS.md / plugins / hooks / skills / MCP, parallel subagents, worktrees, headless -p (Grok Build). Mixing those two up is how you end up comparing a weight file to a CLI product.

For subscription tiers and API cost details, see the full Grok pricing guide.

Full model card and install paths: How to use Grok 4.5 for coding. Agent-vs-agent (not model-vs-model): Grok Build vs Claude Code vs Codex.

Why Grok 4.5 makes remote control more useful, not less

First principles: a coding agent is most valuable when it can grind for a long stretch without you typing every step. That same strength creates the failure mode everyone hits:

  1. You kick off a multi-file task before lunch.
  2. The agent hits an approval prompt four minutes in — or drifts down a path you’d kill in five seconds.
  3. You aren’t at the keyboard. The run freezes, or it burns tokens and wall-clock going the wrong way.

Grok 4.5’s economics make that failure mode more expensive to ignore, not less. At $2 / $6 it is cheap enough to run long loops and frequent retries; at ~80 TPS those loops finish faster, so “I’ll check when I’m back” often means the damage is already done. A model that is fast and thrifty is exactly the kind you want on a leash you can hold from your pocket — watch tool calls, course-correct mid-run, approve or reject without walking back to the desk.

Official first-party paths cover Grok Build on the machine you’re sitting at, and Cursor in the editor. What they don’t give you is a single mobile surface that also runs Claude Code and Codex next to Grok while the agents stay on your hardware. That’s the gap SeaWork is built for.

Run Grok 4.5 in Grok Build (at your desk)

Install Grok Build, then run a task. Shape:

# Official install (or install Grok from SeaWork desktop Settings)
curl -fsSL https://x.ai/cli/install.sh | bash

# Interactive Grok Build session (Grok 4.5 is the default model post-launch)
grok "refactor the auth layer and keep the test suite green"

API path if you want the model without the agent shell:

curl -s https://api.x.ai/v1/responses \
  -H "Authorization: Bearer $XAI_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "grok-4.5",
    "input": "Find and fix the bug, then explain it: function median(a){a.sort();return a[a.length/2]}"
  }'

(SpaceXAI; docs)

That’s fine when you’re at the desk. The moment you start a long Grok loop and want to walk away, you want remote control of the same local agent, not a second copy of the repo in someone else’s cloud.

Steer a Grok Build agent from your phone with SeaWork

Don’t move the agent. Move the interface.

Grok Build on your laptop is already wired into your world — credentials, MCP servers, real files in the real repo, real tool versions. Recreating that in a hosted sandbox is the expensive path. The clean pattern is: leave the agent where the environment already works, and bring the control surface to your phone.

SeaWork does that. The daemon runs on your development machine and drives the agents already installed there — including Grok (Grok Build via ACP). Your phone connects over an end-to-end encrypted relay that only forwards encrypted packets; it never holds the keys to your prompts, diffs, or logs. Code, logs, and provider credentials stay on your machine. Prefer no relay? Self-host the daemon on your own network.

You don’t install a separate daemon — the desktop app ships with it. Three steps:

  1. Install the desktop app. The SeaWork daemon is built in. If Grok Build isn’t on the machine yet, install it from desktop Settings (or the official CLI install above).
  2. Pair your phone. Open the mobile app and scan the pairing code to establish the encrypted link.
  3. Start a Grok agent from the phone, the desktop, or the terminal — default model is Grok 4.5.

CLI shape under the hood:

# Start a Grok Build agent on a task (Grok 4.5 in your setup)
seawork run --provider grok "refactor the auth layer"

# List agents (running and finished)
seawork ls -a

# Follow logs live
seawork logs <id> -f

Start something like that, then pick it up on your phone. You see every tool call as it happens, read the full conversation, and message the agent mid-run when it drifts — before another twenty minutes of thrift-priced tokens go the wrong way. When it hits an approval prompt — the gate that freezes a long run until you’re back at the keyboard — the request lands on your phone and you approve or reject in seconds.

That’s “Grok remote”: not a different Grok product, and not a cloud clone of your monorepo. It’s the same local Grok Build process, with the control surface in your pocket.

When Grok 4.5 is the right remote agent (and when it isn’t)

Match the model to the task, not the launch post.

Reach for Grok 4.5 + remote control when:

  • You want frontier-ish coding at ~$2 / $6 and expect long agent loops
  • Speed and token thrift matter — volume work, iterative refactors, agent fan-out
  • You’re already on Grok Build / SuperGrok and want that default model on a phone leash
  • You mix agents and need Grok beside Claude Code and Codex in one app

Reach for something else when:

  • The task is hard, repo-level engineering where SWE-Bench Pro-style difficulty dominates — Claude Fable 5 still leads that axis; see Grok 4.5 vs Claude Fable 5 vs GPT-5.6
  • You need a three-tier cost ladder inside one vendor family — GPT-5.6 Sol / Terra / Luna is built for that; see GPT-5.6 with Codex from your phone
  • You need EU availability right now — Grok 4.5 was not in the EU at launch; check SpaceXAI’s current status before you plan on it

Most working engineers will not pick one agent forever. They’ll run Grok for volume and speed, Fable 5 for the gnarly one-pass refactor, and Codex when the GPT tier ladder fits the budget — ideally from one remote surface so the desk isn’t a single point of failure.

One app for Grok, Claude Code, and Codex

Grok 4.5’s price changes the economics of mixing models. You might run Grok on the bulk of a feature, hand Fable 5 the one hard subtask, and keep a Codex Terra agent on tests in the same afternoon.

SeaWork runs Grok, Claude Code, and Codex from a single app, side by side, and in parallel: one agent on the refactor, another writing tests, a third drafting migration notes — each in its own workspace, each visible at a glance so you see what’s running, what’s waiting on you, and what finished.

For anyone splitting implementation, tests, and review across worktrees — the workflow Grok Build’s parallel subagents and Claude/Codex multi-agent patterns both encourage — a single terminal scrollback stops scaling. A phone that shows every agent at once fits that far better.

Wider surveys of the same pattern:

FAQ

Does Grok have official remote control like Claude Code /rc? Not as a first-party “Remote Control” product in the same shape as Claude’s /rc. The practical remote path for a local Grok Build session is a control surface that attaches to the machine where Grok is already running — which is what SeaWork does.

Is this Grok chat on mobile? No. Grok’s consumer chat apps are a different product. This page is about Grok Build / Grok 4.5 as a coding agent on your machine, steered from your phone.

Does my code leave my machine? With SeaWork, the agent process stays local. The relay only forwards end-to-end encrypted packets; it never has the keys to read prompts, diffs, or logs. You can also self-host without depending on the public relay.

What provider id do I use in the CLI? grok — e.g. seawork run --provider grok "…". Default model in SeaWork’s Grok provider is Grok 4.5.

Can I use Grok 4.5 in the EU? Not at the July 8, 2026 launch. SpaceXAI expected mid-July availability; verify on x.ai/news/grok-4-5 and the console before you depend on it.

Where to start

Using Grok 4.5 in Grok Build at your desk is the easy part — install, default model, go. The more valuable part is not being pinned to that desk while a long, thrifty Grok loop grinds through a multi-file task. That’s exactly when watching and steering from your phone turns frozen approvals and wasted tokens into a two-second course-correction.

SeaWork is free. Download SeaWork, browse models available in SeaWork, or start from the pure-content deep dives if you still need the model math: Grok 4.5 for coding, Grok 4.5 vs Fable 5 vs GPT-5.6, Grok Build vs Claude Code vs Codex.

Sources