Aeon vs Claude Duet: the agent that asks, the one that runs

On 21 October Anthropic opened a research preview of Claude Duet, "a new app where Claude works as a proactive agent." It sends you a morning brief, watches your connected apps through the day, takes on tasks by message, and keeps working after you close the app. Aeon has been doing a version of the agent that runs while you sleep for a while, but it comes at the problem from the other end.

This is a comparison built from Anthropic's published material (the announcement, the help-center guide and the safety post from 23 October) and Aeon's docs. Duet is waitlisted, so this is not a hands-on review. Disclosure up front: I contribute to Aeon, an open-source agent framework created by Aaron Elijah Mars.

Claude Duet is a hosted personal agent for your inbox, calendar and business apps. It runs on Anthropic's cloud and asks before anything consequential. Aeon is an open-source agent that lives in a GitHub repo you own and runs skills on a cron through GitHub Actions. It is built to run unattended. They overlap less than the "always-on agent" label suggests.

From Anthropic's own pages:

• Who it's for. "Professionals and small business owners with more on their plates than they can track."
• What it does. A morning brief across your connected apps, then check-ins "every few hours" with a message only when something needs you. You can also hand it larger jobs like a dashboard, a financial model or a deck, and the results land in a Make tab.
• Where it runs. A virtual machine in the cloud with its own browser, so it "doesn't require your laptop or phone to be on."
• Model. Claude Opus 5.5. You can't pick a different one, and it decides how much effort a job needs.
• Connectors. It carries over the memory, skills and connected apps from your Claude account, and can use the 6,000+ connectors and plugins in the Claude Marketplace. Gmail, Calendar, Slack, QuickBooks, Stripe, Gusto, HubSpot, Docusign and Shopify are the examples given.
• Approvals. Before it sends a message, makes a payment, signs a document or deletes anything, it shows you the action and waits. Trading securities and moving money between accounts are off-limits even with approval.
• Access and cost. A waitlisted research preview for US Pro and Max subscribers, on iOS, Android and claude.ai. It's included in the plan, and its background checks count toward your usage limits.

From the Aeon docs:

• Who it's for. Anyone who wants recurring agent work done without supervision: research digests, PR review, market and on-chain monitors, content, security scanning. The public catalog lists 86 skills.
• Where it runs. Your own GitHub repository, on your own GitHub Actions. A scheduler workflow ticks every five minutes by default and dispatches any skill whose cron is due. There is no Aeon-hosted runtime: Aeon Connect provides the dashboard and an MCP server, and the docs say it "stores no model keys."
• Model. Per-skill. The docs suggest Sonnet 5.5 as the default, Opus 5.5 for deep reasoning and security audits, and Haiku 5.5 for scoring. Skills can also run through nine coding-agent CLIs, including Claude Code, Codex, Grok, Cursor and Hermes.
• Memory. Plain Markdown and JSON in a memory/ folder, read at the start of each run and committed back to git at the end.
• Approvals. The tagline is "no approval loops, no babysitting." The safety work goes into the structure instead (more on that below).
• Access and cost. MIT licensed. A public repo on a free GitHub account with an existing Claude Pro or Max subscription costs $0 in extra infrastructure. A private repo draws on your Actions minutes (2,000 a month on GitHub Free).

Ownership and visibility

Duet's state lives in Anthropic's product. You can see its task list, steer it mid-task, and reset its memory from settings. Aeon's state is a set of files in a repo you control. Every run is a GitHub Actions log, and every memory change is a commit you can read, diff or revert. If you want to know why the agent did something last Tuesday, with Aeon the answer is in your git history.

That cuts both ways. Owning the runtime also means owning its upkeep. Aeon handles some of that itself: skills are scored 1 to 5 after each run, and a self-healing loop files issues and opens fix PRs. Even so, it's your repo to look after. Duet asks nothing of you beyond joining the waitlist.

Approvals versus unattended runs

This is the real design split. Duet is built to stop and ask. Anthropic's safety post spends real space on calibrating it, noting that "approval fatigue is a serious issue, because too many requests teach people to tap through them." A separate classifier, built on Claude Code's auto mode classifier, checks each action against your intent, and it runs outside the model doing the work.

Aeon assumes the job is something you want done on a schedule without anyone watching, and it puts the limits into the run itself:

• Skills marked mode: read-only run inside a Linux bwrap read-only filesystem, with write, git and gh tools removed.
• A Bash analyzer blocks any command that contains a bare $SECRET. Authenticated calls go through ./secretcurl with {ENV_NAME} placeholders, and only secrets the skill declared up front get substituted.
• Deploys, payments and on-chain transactions go through ./secretcurl as the skill's final step, and they fail closed.
• An optional Fleet Watcher can consult an outside authorization server before every run and block it.

Neither is right in general. For an agent that sends email as you, signs contracts or buys things, a human approval step is the right default, and Duet's design reflects that. For a weekly vulnerability scan or a daily research digest, an approval per run would defeat the point of scheduling it. The longer version of Aeon's approach is in least privilege for an autonomous agent.

Credentials

Both keep secrets away from the model, but at different strengths. Duet "never sees your passwords, browser cookies, sign-in tokens, or other credentials." Passwords are encrypted on your device and opened only on a separate browser machine. Connector tokens follow the Claude Managed Agents design, where a proxy outside the agent's machine swaps in the real value as a request goes out. Aeon keeps secrets in GitHub's encrypted Actions secrets and uses placeholder substitution through ./secretcurl, so the model writes {ENV_NAME} instead of handling the value. The idea is similar. Duet's version adds separate machines, which is stronger isolation and fits an agent that logs in to your personal accounts.

Cadence and cost

Duet checks your accounts at set intervals, and Anthropic says why: "checking nonstop would use up your plan's limits much faster." Anthropic's own docs also note that something urgent can arrive between checks. Aeon's cadence is whatever cron you write for each skill. The trade-off moves elsewhere: on a private repo, frequent runs spend Actions minutes, and the docs suggest widening the scheduler tick if that adds up.

Who can use it today

Duet is US-only for now, Pro and Max only, invitation by waves, with Team and Enterprise "not included yet." Aeon can be cloned or launched from the browser by anyone with a GitHub account and a model credential.

Duet is a polished product with a mobile app, voice, one ongoing thread and a large connector catalog. Its payment flow uses one-time cards from Stripe's Link or Shop Pay, limited to one merchant and amount, so your real card number is never shared. It has a classifier that blocks sharing credentials, card numbers and government IDs even if you ask. An external firm penetration-tested it with access to the source code, and it's in scope of Anthropic's public bug bounty. For a small-business owner who wants help with invoices, payroll and an overflowing inbox, that is a lot of finished work Aeon doesn't try to do.

Aeon is aimed at recurring work that should run on schedule whether or not anyone is looking: scanning code, reviewing PRs, tracking markets, writing digests. Its vuln-scanner skill is a good example. It runs Semgrep, TruffleHog over git history and osv-scanner on a cron, then has the model triage the results down to exploitable findings. Aeon's security page lists the disclosed results. Because it's open source and lives in your repo, you can read every skill prompt, fork it, change its model or harness, and see exactly what ran.

They're not mutually exclusive. Aeon can authenticate with a Claude Pro or Max subscription token. Its MCP server can be added to Claude, which lets you run Aeon skills and read its memory from a chat. Anthropic says Duet carries over the connected apps from your Claude account. Whether that includes a custom MCP connector like Aeon's isn't documented, and it hasn't been tested yet, so treat it as an open question rather than a feature.

• Your day is email, calendar, invoices, payroll and follow-ups, and you want help with all of it from your phone.
• You want an approval step before anything goes out in your name.
• You're in the US on a Pro or Max plan and don't mind a waitlist.

• The work is recurring and should run unattended on a schedule: scans, reviews, digests, monitors.
• You want to own the runtime, read every prompt, and see every run and memory change in git.
• You want to choose the model per job, or you're outside the US.

Is Claude Duet free?

During the research preview it's included in Pro and Max at no extra cost. It uses your plan's usage limits, and its background checks and morning brief count toward them.

Can I use Claude Duet outside the US?

Not yet. The preview is for US Pro and Max subscribers only, and Team and Enterprise plans aren't included.

Is there an open-source alternative to Claude Duet?

Aeon is open source under the MIT license and also keeps working after you close the laptop. It's aimed at scheduled work rather than your inbox, so it's an alternative for some jobs, not a clone.

Can Aeon use Claude models?

Yes. The default harness is Claude Code, authenticated with a Claude Pro or Max subscription token or an Anthropic API key, and you can pick Sonnet, Opus or Haiku per skill.

Can Claude Duet spend money?

Yes, with your approval. It pays with a one-time card from Stripe's Link or Shop Pay, limited to one merchant and amount. Trading securities and moving money between accounts are off-limits even if you approve. Aeon's take on agent spending is in letting your agent spend money.

Can I run both?

Yes. They don't share a runtime, and Aeon can run on the same Claude subscription. See "Using both" above for the one open question.

• Aeon vs OpenAI Dots
• Aeon vs Meta Muse
• Aeon vs Grok Bot
• Aeon vs Claude Code

• Anthropic, Introducing Claude Duet (21 October 2026)
• Anthropic, Claude Duet product page
• Anthropic, Get started with Claude Duet
• Anthropic, Claude Duet Safety, Privacy, and Security (23 October 2026)
• Aeon docs: Get Started, Security, Cost, Scheduling, Memory, Harnesses
• Aeon, Aeon MCP Server: Drive Your Agent from ChatGPT or Claude (20 October 2026)