Amp Orbs Desktop Control — The No-Subscription Agent Just Got Eyes

Amp shipped “See and control the desktop in your orb” on September 4, 2026. That single changelog line means agents running inside Amp’s remote Orb environments can now interact with a high-resolution interactive Linux desktop — browsers, Electron apps, Android emulators, build UIs — without touching your local machine. Nine days later, on September 13, Amp made BYOK (Bring Your Own Key) model routing generally available, letting teams plug their own Anthropic or OpenAI keys into Amp with zero markup on provider API costs. The combination makes Amp distinct among major public coding agents: remote desktop-capable Orbs paired with BYOK, zero-markup pricing and no flat subscription requirement.

TL;DR

• What: Amp Orbs now include an interactive Linux desktop for agentic GUI interaction — released September 4, 2026
• Pricing: BYOK model routing (September 13) means zero Amp markup on provider calls; you pay API cost + Orb compute hours, nothing else
• Contrast: Claude Code requires a $20+/month subscription; Amp charges usage only
• Caveat: Computer use in Orbs is explicitly “early and unsupported” — Amp’s own docs say so

What Happened

The Desktop tab in any Orb thread now exposes a full interactive Linux desktop. This is not a screenshot preview or a static render — the agent can see and manipulate GUI elements inside the remote VM that already runs your repository. If you are building a Chrome extension, an Android app, or anything that requires visual verification, the agent can now open the compiled output, interact with it, and verify the result without your laptop being involved.

The same week, Amp released a Model Routing panel that lets teams independently configure which provider handles each agent role — Oracle (high-reasoning), standard agent, and autocomplete — using their own API keys. For individuals and non-enterprise workspaces, Amp applies no markup to provider API prices. Your bill is provider cost plus Orb compute hours. That is the entire invoice.

# Amp pricing tiers (September 2026)
# Hobby:     Free — BYOK only, bring your own model keys
# Megawatt:  $20/month — $20 included usage, 750 hours of orbs, ~60% discount on additional usage
# Gigawatt:  $200/month — $200 included usage, 1,000 hours of xxlarge orbs, ~65% discount

Amp’s own documentation labels computer use in the Orb desktop as “early and unsupported.” This is a capability you can test today, not one you should wire into production CI/CD pipelines tomorrow.

Why This Matters

Every other major coding agent offering desktop-level computer use runs it through a subscription gate. Claude Code requires at least a Pro subscription at $20/month before you can touch computer use, and it runs on your local machine or Anthropic’s managed runner — both models where the agent operates inside your trust boundary or Anthropic’s. Amp Orbs flip that: the agent runs in a remote VM that Amp manages, and desktop control happens inside that remote environment. Your local machine stays clean.

The architectural difference is not cosmetic. When Claude Code does computer use, it needs access to your desktop — your browser tabs, your file system, your running processes. The attack surface scales with everything on your machine. With Amp Orbs, the attack surface is bounded by what is inside the Orb: your repository clone, the Linux desktop environment, and whatever the agent installs there. Nothing on your local machine is exposed.

The BYOK pricing model amplifies this. A team running Claude Code for agentic work pays Anthropic’s subscription fee regardless of usage. A team running Amp with BYOK pays only when the agent does work, at the exact price the model provider charges. At low-to-moderate usage — say, 30-40 agentic sessions per month — the BYOK model undercuts a $20/month flat seat handily. At heavy usage, the math flips, and Amp’s Megawatt ($20/month with $20 included usage and roughly 60% discount on additional usage) or Gigawatt ($200/month with 65% discount) tiers become the better deal. Either way, the pricing is legible. You know exactly what you are paying for.

This matters because the subscription-to-metered shift is happening across the industry. GitHub Copilot moved to token-metered AI Credits in June 2026. Anthropic split Agent SDK billing into separate credit pools the same month. Amp skipped the awkward transition entirely — they paused their free tier when frontier tokens got expensive, admitted it publicly, and rebuilt around BYOK. That is honest pricing even if the business model sustainability remains unproven.

If you are evaluating Amp Orbs for QA automation — verifying that a built UI renders correctly, that a Chrome extension loads, that a generated Android layout does not break — this is the strongest use case today. Greenfield development inside Orbs still depends on benchmark parity with Claude Code and Codex, where Amp trails on published leaderboards.

The competitive positioning tells you where to use this and where not to. Amp’s code-graph retrieval layer — originally built during the Sourcegraph era before Amp became independent in December 2025 — provides deep retrieval across large repositories that frontier-model agents doing naive file search cannot match. Desktop computer use in Orbs is additive to that retrieval advantage. But on raw coding benchmarks, Amp is not leading the pack. Claude Code and Codex consistently top published leaderboards. If your workflow is “write new code from scratch as fast as possible,” Amp is not the answer. If your workflow is “navigate a large existing codebase, make targeted changes, and visually verify the result in a remote environment without burning my local machine,” Amp just became significantly more capable.

The Take

I have been skeptical of the pay-as-you-go agent story because every tool that started there eventually flipped to subscriptions once frontier token costs hit. Gemini CLI, the old Amp Free tier, half of 2026’s disruption plays — all followed the same arc. But the Amp Orb model is structurally different in a way I did not expect: the remote execution environment and the model call are billed together on one invoice, with no Amp markup, using your own provider keys. That is not a pricing gimmick. That is a different architecture.

The real question is survival. Amp already killed its free tier because frontier tokens got expensive faster than revenue followed. They do not intend to bring it back. A company that openly admits its previous model was unsustainable gets more trust from me than one that quietly jacks up prices — but trust does not pay server bills. If Amp cannot grow its paid base fast enough, the BYOK model dies with it.

One governance gap worth flagging: BYOK means zero markup on provider API calls, but Amp still routes every request through its infrastructure. Zero-markup pricing does not equal zero data exposure. Amp claims a zero-and-minimal data retention policy extended to all users and workspaces, but the specific retention window for Orb session logs — what the agent saw on your desktop, what files it touched, what it rendered — is not publicly documented. Before connecting a production repository to an Orb with desktop control enabled, verify what Amp retains and for how long. The pricing transparency does not automatically extend to the execution trace.

I would use this for QA automation before I would use it for greenfield development. And I would use it with a repo I can afford to have routed through third-party infrastructure. Under those constraints, Amp Orbs with desktop control is the most cost-transparent remote agent environment available right now.