DigitalOcean Cuttlefish Review: Approve AI Agents From Your iPhone

Your coding agent has been refactoring for two hours. It is deep in the codebase, making real progress, and then it stops: “Can I run this migration?” Nothing happens until you answer, and you are not at your desk. You are in line at the grocery store, or waiting at school pickup.

That is the exact moment DigitalOcean built Cuttlefish for. Launched on November 5, 2026 at BoxWorks, Cuttlefish is a native iPhone app for DigitalOcean Managed Agents that lets you follow a live agent session, approve or deny its tool calls, and steer or stop the run from your phone. The agents themselves run in the cloud, so the work keeps going after you close your laptop. Now the human in the loop can be anywhere too.

It is a small app for a real bottleneck. Agents that run for hours stall on the small decisions only a person can make, and every stalled hour is a wasted hour of cloud compute. Cuttlefish does not make your agent smarter. It makes the approval step fit in your pocket. This review covers what the app does, how to set it up, what the agents behind it cost, and the limits in DigitalOcean’s own documentation.

What Cuttlefish actually is

Cuttlefish is not a coding agent. It is a remote control for agents that already run on DigitalOcean Managed Agents, the company’s hosted platform for running coding agents in isolated cloud sandboxes. Managed Agents has been in public preview since September 2026, and each session runs in its own hardware-isolated microVM, a lightweight virtual machine that persists independently of any device. Close your laptop, lock your phone, or close the app: the session keeps running on DigitalOcean’s side.

The app supervises whichever harness your team runs on Managed Agents, including Claude Code, Codex CLI, OpenCode, OpenHands, Hermes, LangGraph, and CrewAI. That matters because the app is not tied to one vendor’s agent: if your team switches harnesses, the same phone app still works. DigitalOcean says the platform gives agents access to more than 20,000 tools, and the app is the window into all of them.

There is one hard requirement that will rule some readers out immediately: Cuttlefish needs an iPhone running iOS 26.5 or later, and a DigitalOcean account on a team with access to Managed Agents. There is no Android version and no iPad version announced. If your phone is older than iOS 26.5, this app is not for you yet.

What you can do from your phone

The app covers four jobs, and it does them with more detail than a notification shade ever could. First, follow live sessions. Replies, the agent’s reasoning, and tool calls stream in as the agent works, and you get notified when a run needs approval, finishes, or fails. Cuttlefish caches transcripts on the device, so reopening a session picks up where you left off instead of replaying its history. A run you stop yourself is not treated as a failure, so your history stays honest.

Second, approve or deny tool calls. When an environment’s permission policy is set to ask for an action, the agent waits for your decision. Cuttlefish shows the request with the full command, file path, or diff, not a summary. The app flags requests it considers risky and says why: a recursive delete, a force-push, a command run with sudo, or a script piped from the internet into a shell. Allowing a risky request requires Face ID or your device passcode. That is the right call for a phone app, because a pocket approval is also a pocket accident waiting to happen.

Third, steer mid-run. You can send a correction while the agent is mid-run, pause, resume, or stop a session that is heading the wrong way. Messages you send while the agent works get queued, so you can line up the next task without waiting. You can even dictate messages: dictation uses on-device speech recognition, and the audio is not sent to DigitalOcean.

Fourth, work with the results. You can view what your agent built, including websites and documents, right in the app. When you open Artifacts, Cuttlefish scans the sandbox for open ports serving web content, such as a Next.js or Python development server, and lists them so you can tap a server to view it. You can browse workspace files, upload files and photos, create checkpoints, and fork a session from a checkpoint or roll one back.

One honest gap: notifications arrive promptly while Cuttlefish is open, but in the background iOS decides when the app checks for changes, so they can arrive many minutes late, and they do not arrive at all after you force-quit the app. If your approval is the only thing standing between an agent and a deploy, do not force-quit the app and walk away.

How to set it up

Setup is short, but the sign-in step has a trap worth knowing about before you start.

Create a DigitalOcean account if you do not have one, and make sure your team has access to Managed Agents. The app is available to Managed Agents customers only.
Download Cuttlefish from the App Store and open it.
Tap Sign in with DigitalOcean. Sign in in the browser sheet that opens and choose the team you want to use.
Approve every OAuth scope the app requests. Cuttlefish asks for session read, create, operate, and manage scopes, plus Action Gateway read and inference read. If you clear any of them, sign-in fails with “Cuttlefish needs access to Managed Agents to work.” If you pick a team that cannot use Managed Agents, sign-in fails and asks you to pick another team.
Start a session by choosing your coding agent, Claude Code, OpenCode, or Codex CLI for example, or open an existing session from the home screen.

Each sign-in opens a browser with no saved cookies, so it always asks which account to use. To add more teams, or to switch between them, use the settings menu in the sidebar. Cuttlefish also includes a guided form that builds an environment spec for you. Secrets you enter in the form go into DigitalOcean’s secret store, are removed from the spec, and are never shown again. To let the agent clone, commit, and push with your team’s GitHub account, turn on Allow GitHub access in the form and tap Connect GitHub.

What it costs: the app is free, the agents are not

Cuttlefish itself costs nothing. The bill comes from the Managed Agents sessions it supervises, and that billing is metered in small pieces. DigitalOcean’s own Harness Runtime pricing page, last verified November 5, 2026, breaks it down.

Compute is billed per second of active CPU, based on actual consumption. When an agent is waiting and consuming no CPU, its CPU charge falls to zero. In DigitalOcean’s own phrasing, billing follows active CPU use, so agent wait time has minimal cost. Memory is billed on allocation over time, and session storage on capacity held. Pausing a session stops compute charges while preserving the sandbox’s files and conversation history.

Sandbox size	vCPUs	Memory	Price at full allocation
XSmall	1	1 GB	$0.0535/hr
Small	2	2 GB	$0.107/hr
Medium (default)	2	4 GB	$0.126/hr
Large	4	8 GB	$0.252/hr
XLarge	16	32 GB	$1.008/hr

Underneath those shapes, the component rates are $0.044 per vCPU-hour, $0.0095 per GB-hour of memory, $0.05 per GiB-month of session storage, and $0.01 per GiB of public internet egress. Snapshots and checkpoints cost $0.05 per GiB-month as well, and they keep accruing storage charges while a session is paused, including when all your balances are empty. Pausing stops compute, not storage.

Tool calls go through Action Gateway, which charges $0.10 per 1,000 standard tool invocations. Tool search is included at no extra charge. Paid tools cost more: Exa web search is $10.10 per 1,000 calls and Exa web fetch is $3.10 per 1,000 calls, with the gateway charge already included. Model inference is billed separately at DigitalOcean’s inference rates for DigitalOcean-hosted models, and directly by the provider when you use your own external keys.

Then there is the subscription layer. Agent Droplets are monthly plans that discount your Managed Agents usage, also in public preview. They add no compute, models, or tools of their own. Each month your plan fee funds a subscription balance, and your usage draws from it at the plan’s discount until it runs out.

Plan	Monthly fee	Discount	Usage covered at list price
Pro	$50 (~£37.15 / ~C$70.08)	15%	$58.82
Team	$200 (~£148.60 / ~C$280.34)	20%	$250.00

Plan fees are billed in advance, exclude taxes, and renew automatically until you cancel. New customers get a $5 promotional credit (about £3.71 / C$7.01). DigitalOcean’s own estimates say that $5 covers roughly 3.5 million to 5.5 million tokens and 60 to 200 agent hours depending on workload, while the Pro plan’s typical workload is about 50 million tokens, 1,500 agent hours, and 50 GB of storage. Those are estimates, not an included allowance.

The discount has a catch worth reading twice: it applies to compute, memory, storage, inference on DigitalOcean-hosted models such as Kimi K3 and GLM-5.3, and Action Gateway tools. It does not apply to third-party models such as Anthropic Claude and OpenAI GPT, including when you use them through DigitalOcean’s serverless inference. Third-party model usage still draws from your subscription balance, but at standard rates. If your agents mostly run Claude or GPT, the plan discount does far less for you than the headline suggests.

What does a real run cost? One independent tester ran five issue-to-pull-request attempts with OpenCode on Managed Agents in September 2026 and published the invoice line by line. The five runs were billed $0.32 in total, about $0.064 per run, or roughly 11 cents per merged pull request, on DeepSeek V4 Pro at DigitalOcean’s list price of $1.74 per million input tokens and $3.48 per million output tokens. A separate 39-minute interactive session on GLM-5.3 cost $0.78, almost all of it tokens. Those are one tester’s numbers on cheap hosted models, not a guarantee, but they show the shape of the bill: tokens dominate and a short coding task costs cents. The same invoice showed a quirk to budget for: every token line was rounded to the cent, with a one-cent minimum, which added 17% to the five runs.

There is no per-session or per-product spend limit on Harness Runtime or Action Gateway. DigitalOcean’s own docs warn that a looping or compromised agent using tools that require prepayment can drain your balances quickly. When all your balances are empty, the gateway blocks calls to tools that require prepayment until you add funds, but tools that do not require prepayment are not blocked. Set your permission policies before you hand an agent a long leash.

Box Mount: your files, your agent, your phone

Cuttlefish launched alongside a collaboration with Box, announced at the same BoxWorks event. Box Mount, now in private preview, syncs a Box folder into your Managed Agents environment, so agents work on the same files your team already uses, under Box’s permissions, version history, and audit trail. In Cuttlefish, you open Customize, then Tools, then Box, sign in, and pick a folder. The agent works on plain files in its workspace, its edits upload to Box, and changes made in Box show up in the workspace. Each person syncs with their own Box account. DigitalOcean’s launch example is a legal team whose agent stops to ask permission before writing a summary of vendor contracts, a lawyer approves from their phone, and the summary lands back in the shared folder.

Two things to note before you get excited. First, Box Mount is in private preview, so most readers cannot use it yet. Second, the agent can edit and delete files in the synced folder, and deletions sync back to Box. That is the point of the feature, but it is also exactly the kind of action you want behind an approval gate, not an auto-approve. Box also announced its own agent tools at BoxWorks, including Box Context Retrieval and Box Architect, plus a partnership with NVIDIA to bring Nemotron retrieval models into Box.

What it cannot do yet

DigitalOcean’s own documentation lists the limits plainly. You cannot create or edit Toolbelts in Cuttlefish, edit the tools on an existing environment, or connect provider accounts for Action Gateway tools. The app gives you aYou cannot answer a question the agent asks: if the agent asks something rather than requesting a tool call, you answer it from doctl or dismiss it in Cuttlefish. And you cannot edit code or do in-depth code review in the app. Cuttlefish is a supervision console, not an IDE, and reviewing a diff on a phone screen is a check, not a review.

The bigger caveat is in the press release’s fine print: Managed Agents is offered as a public preview, and Box Mount as a private preview. Preview services are provided “as is”, are not supported for production workloads, and may change, never become generally available, or be discontinued. That is standard preview language, but it means the pricing, the shapes, and the features above are all subject to change.

There is also no independent testing of the app itself yet. Everything in this review comes from DigitalOcean’s launch blog, its documentation, the press release, and one independent invoice analysis, not from hands-on use. The approval flow, the Face ID step, and the notification timing are DigitalOcean’s descriptions, and the first real test will be whether approvals arrive fast enough when an agent is waiting on a deploy at 2 a.m.

Cuttlefish vs the other ways to supervise an agent

Approving an agent’s work from your phone is not a new idea, but the options differ in where the agent lives. Cuttlefish runs the agent in the cloud and puts the approval step on your phone. A laptop terminal runs the agent locally and gives you full control, until you close the lid or lose Wi-Fi, at which point the session stops with the machine. OpenClaw lets you text an assistant from WhatsApp or Telegram, but the assistant runs on your own machine, so the same laptop problem applies.

Option	Where the agent runs	Approve from anywhere	What you give up
Cuttlefish + Managed Agents	Cloud microVM	Yes, iPhone app	Code editing, in-depth review, Android
Laptop terminal (Claude Code, OpenCode)	Your machine	No, tied to the laptop	Nothing, but the session dies with the laptop
OpenClaw via WhatsApp	Your machine	Yes, by text message	Cloud persistence, visual diffs, checkpoints

If you already run Claude Code or OpenCode on your laptop, Cuttlefish is the upgrade path for the moments you walk away from it: the same harnesses, the same sessions, now running in a sandbox that does not care whether your laptop is open. If you have never run an agent in the cloud, the $5 trial credit is the honest way to find out whether the meter fits your habits before the $50 plan does.

Pro tips and common mistakes

Set the permission policy before you leave your desk. Cuttlefish only asks you about actions your environment is configured to ask about. If everything is auto-approved, the app has nothing to approve, and the safety story collapses. Decide which actions need a human before the agent starts, not after it has already run a migration.

Pause sessions you are not using, then check your checkpoints. Pausing stops compute charges, which is the meter that matters most. But checkpoints and snapshots keep accruing $0.05 per GiB-month while a session is paused, even when your balances are empty. Delete stale checkpoints, or the paused session keeps billing you in storage.

Do not force-quit the app. Background notifications can already arrive minutes late because iOS decides when Cuttlefish checks for changes. After a force-quit, they do not arrive at all. If you are the approver on a live run, leave the app alone.

Treat Box Mount folders as live. The agent can edit and delete files in a synced Box folder, and deletions sync back. Keep the approval gate on for write actions, and do not point it at a folder you would not want an agent to rewrite.

Frequently asked questions

Is Cuttlefish free to download?

Yes, the app is free on the App Store, and there is no charge for the app itself. You pay for the Managed Agents sessions it supervises: compute from $0.0535 an hour for the smallest sandbox, tool calls at $0.10 per 1,000, and model tokens on top. New customers get a $5 promotional credit to try it.

Does Cuttlefish work on Android?

No. Cuttlefish is an iPhone app only, and it needs iOS 26.5 or later. DigitalOcean has not announced an Android or iPad version. Android users can still use Managed Agents from the control panel, doctl, or the API, just not from this app.

Which agents can Cuttlefish supervise?

Any harness your team runs on Managed Agents, including Claude Code, Codex CLI, OpenCode, OpenHands, Hermes, LangGraph, and CrewAI. The app supervises the session, not the harness, so switching harnesses does not mean switching apps.

What happens if my agent loops and burns through my balance?

There is no per-session spend limit, so a looping agent can drain your balances. DigitalOcean’s docs say the gateway blocks calls to tools that require prepayment once all balances are empty, but tools that do not require prepayment are not blocked. Pause the session from the app and set permission policies to ask before expensive actions.

Can I approve a risky command from my phone?

Yes, and that is the point. Cuttlefish shows the full command, file path, or diff, flags risky requests with a reason, and requires Face ID or your device passcode before it lets a risky request through. It is a deliberate extra step, not a tap-through.

Is my data private in Cuttlefish?

Sign-in tokens are stored in your iPhone’s Keychain only and are not synced to your other devices. Cached transcripts and session lists are excluded from device backups and deleted when you sign out. The app includes no analytics, advertising, or crash-reporting SDKs. Prompts, files, and photos you send do go to DigitalOcean and to the model provider your environment uses.

The bottom line

Cuttlefish solves a narrow problem well: the approval step is the bottleneck in long agent runs, and this app puts it in your pocket with the detail, full commands, full diffs, Face ID on risky calls, that makes a phone approval defensible. The pricing underneath is metered and honest, with a $5 credit to try it and a $50 plan that covers $58.82 of usage, but the discount does not touch Claude or GPT tokens, and there is no spend limit to catch a runaway agent. It is iPhone-only, it is still in preview, and it cannot edit code or answer your agent’s questions. If you already run coding agents on DigitalOcean and you keep walking away from them mid-run, download it. If you are still deciding where your agents should live, start with the trial credit and watch the meter before you commit.

Download Cuttlefish from the App Store, sign in with your DigitalOcean account, and the next time your agent stops to ask permission, you can answer from wherever you are instead of coming back to a stalled run.

Sources

Currency conversions use November 6, 2026 market rates (1 USD = 0.743 GBP / 1.4017 CAD) and are approximate. Managed Agents and Cuttlefish are in public preview; prices and features may change.

16 min read

6 views

1h ago

What are you looking for?

Share This Page