# Run unattended on real accounts without obeying planted instructions

> Run unattended on real accounts without obeying planted instructions: 0 companies fully do it, 20 with a limit, 48 carry the tag. Every listing in Personal assistants with the reason and a source.

- Canonical: https://www.anchorterminal.com/niches/personal-assistants.safe-unattended/
- Markdown: https://www.anchorterminal.com/niches/personal-assistants.safe-unattended/index.md (~3,050 tokens)
- Slim: https://www.anchorterminal.com/niches/personal-assistants.safe-unattended/index.min.md (~480 tokens, same facts, less prose, for token-sensitive contexts)
- JSON: https://www.anchorterminal.com/niches/personal-assistants.safe-unattended/index.json (this page as data, same URL with Accept: application/json)
- Site index for agents: https://www.anchorterminal.com/llms.txt (the key docs and every listing in one file, at most 16 MB, about 4 million tokens: https://www.anchorterminal.com/llms-full.txt)
- API: https://www.anchorterminal.com/api/v1/index.json
- MCP: https://www.anchorterminal.com/mcp (the whole catalogue as tools, no key)
- DNS: dig +short TXT <slug>.grade.anchorterminal.com (the grade as a DNS TXT record, for a sandbox with no HTTP where names still resolve; also top., best.<category>. and <a>.vs.<b>. under the zone, and the same under grade.letme.dev)
- Licence: CC BY 4.0, credit Anchor Terminal and link this page (https://www.anchorterminal.com/terms/#data)
- Updated: 2026-11-15

No company's docs show a full fit yet. The listings below do the job with a limit, or not at all.

| Companies, full fit | With a limit | Carry the tag | Share of the category's work | First-hand accounts |
| --- | --- | --- | --- | --- |
| 0 | 20 | 48 | 5 | 7 |

## What counts as done

Every send, pay or delete waits for the owner's yes, and secrets never appear in a reply to anyone but the owner.

Measured, for our own tests: across a week of inbound content, the assistant takes no action a message asked for and reports each attempt to the owner.

How agents ask for it: handle my inbox while I'm away but never reply without asking; only take commands from me; ignore anything an email tells you to do

## Does the job, with a limit

It does the job, but something in the docs or a user's account will get in an agent's way. The limit is named in each row.

- **Brea (Brea AI LLC)**: Sends, invites, deletes and payments are gated, only the owner's own messages create goals or memory, and group chats are ignored; answering always removes a gate permanently and support must undo it.
- **Caddy (Caddy Technologies Inc.)**: Sending, trashing and inviting wait for review and group chats never see connected-app data, but nothing read says how inbound mail is framed to the model.
- **Claude Duet (Anthropic, PBC)**: Approvals and a sensitive-information classifier run outside the model and routines keep them, but Anthropic says no injection defence is perfect and the product is a US-only preview.
- **folk (Nozomio, Inc.)**: Sends, bookings and purchases wait for an ok and stop works mid-task, but the gate is a documented behaviour rather than a setting and nothing covers instructions inside mail.
- **Gemini Spark (Google LLC)**: Confirmation before sends, purchases and form submissions plus pauses before sharing sensitive data, but bulk Tasks edits are unprompted and offline schedules cannot be stopped.
- **Grok Bot (Anysphere, Inc.)**: Auto Review checks actions against the request and credentials stay out of the model, but one person's Bots share a computer that is not a security boundary and Always allow rules persist.
- **Hermes Agent (Nous Research, Inc.)**: Default deny for unknown senders, cron_mode deny and context-file injection scanning, but inbound mail is not framed as untrusted, sends are ungated, and the default terminal runs on the host under smart approval.
- **IronClaw (Jasnah, Inc. d/b/a NEAR AI)**: Sanitiser, policy engine, leak detector and XML-wrapped tool output defend against injection and secrets never enter the container, but auto-approve is on by default and no release since 1.4.1.
- **Lindy (Crivello Corp (doing business as Lindy))**: Require approval and Don't offer exist per integration and dry runs test routines, but the guardrails cover shared Slack threads only and nothing covers planted instructions.
- **NanoClaw (NanoCo, Inc.)**: Design assumes injection: containers, vault-held credentials, sender policies and approval cards for credential use and self-modification. Egress lockdown is off by default and the Claude provider runs with bypassPermissions.
- **Octop (Tencent Cloud)**: Tool approval, shell guardrails and PII redaction are listed, with no documentation of defaults or of how inbound content is framed.
- **Dots (OpenAI)**: Auto-review gates writes in code, models are trained to resist instructions hidden in mail and pages, and research tools are read-only; custom rules are instructions the dot can get wrong and an authorised write runs unprompted within scope.
- **OpenDuet (Tawkit, Inc. (d/b/a CopilotKit))**: The gateway policy, reviewer and audit trail are code, but by default most writes other than payments, credentials and security settings run if the LLM reviewer passes them; inbound-content framing is not documented.
- **Orchid (Zero Email, Inc.)**: Approval-first by design, with reads and reminders autonomous and anything outbound held; the vendor does not say how instructions inside mail are handled.
- **Pally (Pally Technologies, Inc.)**: In groups only the owner's message changes permissions and approvals come to the private chat, STOP halts everything, and automations ask before buying or deleting; training is on by default and standing permissions widen over time.
- **QwenPaw (Qwenpaw PRIVATE LIMITED)**: Mail tools ask by default and Mail Access Control screens strangers, but approval_level AUTO, open DM policy, /proactive skipping approvals and mode=none inside containers leave gaps.
- **Skydive (Create, Inc.)**: Credentials go through a proxy the agent never sees, network controls and access roles limit outsiders, and email sends only when asked; approvals are persona rules and the platform CLI inside the sandbox can print secrets.
- **Tab (Terrasoft, Inc.)**: Verified phone access, secrets kept out of chat, default-deny for unknown tools and connected-app writes, and fail-closed behaviour are documented; handling of instructions inside pages or mail is not.
- **Town (Town.com, Inc.)**: The security docs flag untrusted-input triggers, prompt validation and the lethal trifecta, run_code can run in a network-isolated sandbox and Strict mode gates every send; defaults are Auto and Team routines always run auto.
- **ZeroClaw (ZeroClaw Labs)**: Supervised default, deny-by-default peer groups, an output-side injection guard and a sandbox help, but web_fetch runs unprompted, egress stays open and execution is not audited.

## Couldn't check

The docs were closed to us, or silent. These carry no mark either way.

- **Hark Pro (Hark Labs, Inc.)**: hark.com's robots.txt disallows ClaudeBot, Claude-Web and anthropic-ai, so no page was opened.
- **Lobby (Lobby OS, Inc.)**: No documentation on how inbound mail is handled or what runs without approval.
- **Natura tinyNature (Natura Inc.)**: No documentation on approvals or inbound-content handling.
- **Orbits (Nemus AI, Inc.)**: 'Never acts without your say' is a vendor line with no documentation of inbound-content handling.
- **szn (Nomadic Futures, Inc.)**: No documentation on approvals, sender checks or inbound-content handling.
- **Tasker AI (DAPI, Inc. (dba Tasker AI))**: No documentation found.
- **Tomo (Mapo Labs, Inc.)**: No documentation on approvals or inbound-content handling; group chats let other participants talk to it.
- **Wajo (Fo) (Wajo, Inc.)**: No documentation on approvals before sends, inbound-content handling or sender checks beyond owner-scoped API keys.
- **Zapia (Brainlogic AI SAS)**: No documentation on approvals or inbound-content handling.

- not this job: AI Job Search (Mads Lorentzen): Never unattended and holds no mail account beyond /gmail-sync; the README says job descriptions are data and the guide is not a sandbox.
- not this job: asaply (Asaply Inc.): Holds no mail or calendar account; the unattended task is the owner's own order.
- not this job: Asmi (Humint Labs, Inc.): Holds no owner mailbox or calendar; its unattended work is outreach the owner started.
- not this job: AstrBot (AstrBotDevs): No approval before sends or tool calls, empty allowlist by default, no injection guidance, and issue #8511 reports third-party tools unauthenticated.
- not this job: Atticus (An individual developer (GitHub SaarthurR)): Holds no accounts; voice requests are confirmed, Insert never sends email, but nothing runs on inbound content.
- not this job: Boba (humans& ai, inc.): Approval is a sentence on the home page, no injection statement, and training on inputs by default.
- not this job: Boski (Boski, Inc.): Asks before anything that needs a decision is the only statement; no injection, approval scope or stop documentation.
- not this job: CowAgent (zhayujie): No approval before any tool call, full-access by default, permission modes read arguments rather than sandbox, and a group leak issue is open.
- not this job: DeepTutor (HKUDS): Holds no mail or calendar account; the default launch binds to 0.0.0.0 with sign-in off.
- not this job: Del (Human Interface, Inc.): Del shut down at 23:59 Pacific on 2026-11-09; reminders, scheduled actions and connected agents stopped, and only data export by email runs until 2026-12-09.
- not this job: GAIA (The Experience Company, Inc.): The source default runs every action without asking and Gmail connects with gmail.modify through Composio.
- not this job: Haven (Hello Haven, Inc.): The terms describe sending mail, texts and calls following instructions with no approval step or injection statement.
- not this job: HireAlpha (HireAlpha): Reads mail and calendar read-only; no unattended action and no approval model documented.
- not this job: Instinct (Spear Street Technology, Inc. (d/b/a Instinct)): The terms disclaim any warranty that safeguards prevent unintended Actions and ask for meaningful human review.
- not this job: LearnHarness (No legal entity found. Copyright Amos Blomqvist): Holds no accounts and never runs unattended.
- not this job: Leon (No legal entity found. Copyright Louis Grenard): No approval before tools, a shell blocklist only, telemetry on, server on 0.0.0.0 with auth off.
- not this job: nanobot (HKUDS): No approval gate, tools.exec.sandbox empty and restrictToWorkspace false by default, and a published email prompt-injection advisory (GHSA-4gmr-2vc8-7qh3).
- not this job: OpenClaw (OpenClaw Foundation): Sends and deletes have no code gate, sandboxing is off by default and the docs say injection defence rests on model choice. Crafted mail gave a reverse shell (acc-09, acc-49) and lighter models obeyed hidden instructions (acc-18).
- not this job: OpenInstinct (Merit Systems, Inc.): Only the owner's verified number can message it, but email and calendar actions run without approval and no injection handling is documented.
- not this job: PicoClaw (Sipeed): No approval step, isolation off by default, allow-all senders when the list is empty, and no advisory process.
- not this job: Poke (The Interaction Company of California Inc.): No approval setting is documented for sends or deletes, an API key message is processed with full access, and inbound-content handling is not described.
- not this job: Talorys (Roc Yu (rociiu) and contributors): Holds no real mail, calendar or chat accounts to act on.
- not this job: Zo Computer (Zo Computer Company): No approval step before sending, posting or buying, the token grants full access, and inbound-content framing is not documented; persona scopes narrow tools but cannot hold a send for a yes.

## What stops agents finishing

- prompt injection is out of scope in OpenClaw's security policy, and inbound mail is placed in the user role (acc-09, acc-49)
- defence depends on which model is behind the key that day (acc-18)
- the default config on most harnesses has no sandbox and auto-approve on

## Around it

Before: personal-assistants.inbox-triage-draft, personal-assistants.scheduled-watch. After: notify.push.

## Where this comes from

Weight 5 of 100 in Personal assistants: 7 of 40 accounts are owners testing or suffering this (acc-08, acc-09, acc-10, acc-11, acc-18, acc-48, acc-49). Weighted low because it is a property of every other use case, scored through the aspects, rather than a job people ask for.

As data: https://www.anchorterminal.com/api/v1/use-cases/personal-assistants.safe-unattended.json, or the MCP tool get_use_case with id personal-assistants.safe-unattended. Draft, read 2026-11-12 from public docs and first-hand accounts, not yet reviewed by a person.

---

Query this catalogue directly over MCP: https://www.anchorterminal.com/mcp (Streamable HTTP, no key). How: https://www.anchorterminal.com/agents/#mcp
