# Triage the inbox and draft replies for the owner to approve

> Triage the inbox and draft replies for the owner to approve: 4 companies fully do it, 25 with a limit, 49 carry the tag. Every listing in Personal assistants with the reason and a source.

- Canonical: https://www.anchorterminal.com/niches/personal-assistants.inbox-triage-draft/
- Markdown: https://www.anchorterminal.com/niches/personal-assistants.inbox-triage-draft/index.md (~3,100 tokens)
- Slim: https://www.anchorterminal.com/niches/personal-assistants.inbox-triage-draft/index.min.md (~530 tokens, same facts, less prose, for token-sensitive contexts)
- JSON: https://www.anchorterminal.com/niches/personal-assistants.inbox-triage-draft/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

Four or more companies' docs show a full fit.

| Companies, full fit | With a limit | Carry the tag | Share of the category's work | First-hand accounts |
| --- | --- | --- | --- | --- |
| 4 | 25 | 49 | 18 | 9 |

## What counts as done

A ranked list of threads needing action and a draft saved or shown for each, with no message sent or trashed until the owner says yes in the channel.

Measured, for our own tests: a planted instruction in a message is reported, not followed, and the list and drafts come within one run of the mailbox.

How agents ask for it: which emails need a reply; draft a response to this; unsubscribe me from junk; chase the unpaid invoice

## Does the job

The docs show it meets the test above, and nothing we read says otherwise.

- **Brea (Brea AI LLC)**: Reading, searching and saving a draft need no approval; sending, forwarding and deleting do, and text in an email cannot create a goal or add to memory.
- **Dots (OpenAI)**: Auto-review checks every write-labelled tool call before it runs, asking to draft replies grants no permission to send, and a custom rule can require approval per action class. Proactive research tools cannot send or change apps.
- **OpenDuet (Tawkit, Inc. (d/b/a CopilotKit))**: Reads Gmail, sorts what needs you, and a gateway policy with a reviewer gates every changing action; a message to other people always pauses for Refuse, Allow once or Always allow, with an audit row per decision.
- **QwenPaw (Qwenpaw PRIVATE LIMITED)**: Mailbox tools run over IMAP/SMTP with send_message and deletes on the ask policy by default, Mail Access Control holds unknown senders for approval, and the docs tell the agent to treat every body as untrusted input.

## 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.

- **Boba (humans& ai, inc.)**: Flags what matters, drafts replies and asks Want me to send it; no read-only mode, no statement on planted instructions.
- **Boski (Boski, Inc.)**: Finds threads left hanging and drafts replies in your voice, you hit send; which mailbox providers connect is not stated and no approval setting is documented.
- **Caddy (Caddy Technologies Inc.)**: Triage is read-only, drafts are shown before sending and trashing needs review, but Gmail is the only mailbox and nothing is said about instructions planted in mail.
- **Claude Duet (Anthropic, PBC)**: Asks before sending or deleting with an approval card outside the model, but checks accounts at intervals (acc-27), drafts were worse than Dot's (acc-21), triage fell short (acc-22), and Microsoft 365 is unavailable.
- **folk (Nozomio, Inc.)**: An inbox sweep skill needs the connected email, refund and cancellation emails are drafted and the user says send; Gmail is the only provider named and nothing covers planted instructions.
- **GAIA (The Experience Company, Inc.)**: Drafts are saved to Gmail and sent only when you say so, and trash is on request, but HIL_DEFAULT_MODE is always_allow, so nothing is gated in code until the owner changes it.
- **Gemini Spark (Google LLC)**: Drafts, replies and labels with confirmation before sending communications, but chats train models under Keep Activity and the feature is personal Google Accounts only outside the UK and EEA.
- **Grok Bot (Anysphere, Inc.)**: Drafts to approve are the suggested pattern and approval cards offer Allow once, Always allow or Deny with an Auto Review model behind them, but the boundary is set in the request text and Always allow saves a rule.
- **Hermes Agent (Nous Research, Inc.)**: Approvals gate terminal commands only; gmail send in the skill runs with no code gate, so drafting without sending (acc-31) rests on the prompt. Issue #42307 asks for a shell-free mail tool for untrusted input.
- **IronClaw (Jasnah, Inc. d/b/a NEAR AI)**: Gmail actions include create_draft, send_message and trash_message with tool output wrapped and leak-scanned, but tools marked ask run without a prompt while AUTO_APPROVE_DEFAULT_ENABLED is true.
- **Lindy (Crivello Corp (doing business as Lindy))**: Drafts land in the drafts folder and nothing goes out without approval, but guardrails apply in shared Slack threads only, so DMs, iMessage and SMS are unguarded and reads are never guarded.
- **Lobby (Lobby OS, Inc.)**: llms.txt says it predicts multi-step email workflows, drafts replies and follow-ups and the user approves instead of prompting; Gmail only, and nothing but the vendor's pages describes it.
- **NanoClaw (NanoCo, Inc.)**: A Gmail MCP skill reads, searches and sends; vault secrets can require a per-request approval card, which would gate a send in code, but that is off unless configured and send_message itself is not gated.
- **Octop (Tencent Cloud)**: The IMAP mail connector and built-in tool approval exist, but the docs do not say which actions the approval covers or whether it is on by default.
- **OpenClaw (OpenClaw Foundation)**: Gmail hook routes mail to a restricted reader with untrusted-content wrapping, but no code gate stops a send: the message tool crosses conversations by default and exec approvals cover shell only. acc-08 lost 200 emails, acc-12 sent an unapproved rebuttal.
- **OpenInstinct (Merit Systems, Inc.)**: Gmail connects with the gmail.modify scope, but email actions have had no approval step in code since 2 October 2026, so drafts are not held for a yes.
- **Orbits (Nemus AI, Inc.)**: Scans iCloud, Gmail or Outlook for household-relevant mail, proposes events and tasks and drafts replies, with proposed changes shown for approval; not a general inbox triage and vendor claim only.
- **Orchid (Zero Email, Inc.)**: The safety post says drafts stay drafts until you say send and Orchid can read, summarise and prepare on its own; Gmail only, and the claim comes with no setting or docs page behind it.
- **Pally (Pally Technologies, Inc.)**: Drafts are saved for approval and Pally checks before sending anything not already approved, but Inbox triage archives mail automatically once on, and standing permissions can be offered after repeated approvals.
- **Poke (The Interaction Company of California Inc.)**: Reads, searches and drafts mail; release notes mention email draft approval only when fixing approval loops, and no setting says which actions need approval. The Unite.AI review cited in acc-36 found triage buried urgent mail.
- **Skydive (Create, Inc.)**: Gmail and Outlook integrations read, draft and send, and training mode shows artifacts for approval, but 'never send without asking' is a persona guardrail rather than a code gate.
- **Town (Town.com, Inc.)**: Human in the Loop mode runs reads, labels, archiving and drafts immediately and turns sends to others into approval requests; Strict thread mode asks before anything. The default thread mode is Auto, so the owner must set it.
- **Zapia (Brainlogic AI SAS)**: Reads incoming mail, picks messages needing an answer and leaves a draft in the owner's account; approval before a send is implied by drafts in-account but not documented, and the copy is marketing.
- **ZeroClaw (ZeroClaw Labs)**: Supervised autonomy prompts for medium-risk tools through Telegram buttons or a channel token and times out to deny. Whether send_via counts as medium risk is not stated, so the listing's advice to put it in always_ask stands.
- **Zo Computer (Zo Computer Company)**: Gmail and Outlook tools read and draft, and a persona can be limited to read scopes, but no approval step before a send is documented and the token carries every permission.

## 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.
- **Haven (Hello Haven, Inc.)**: The terms say it sends emails following your instructions and connects to email by type; no draft, review or triage flow is documented.
- **Instinct (Spear Street Technology, Inc. (d/b/a Instinct))**: Connected Services are defined generically; no mail flow is described.
- **Natura tinyNature (Natura Inc.)**: Email triage from Gmail is named; drafting replies or holding sends is not described.
- **Tasker AI (DAPI, Inc. (dba Tasker AI))**: No documentation found.
- **Tomo (Mapo Labs, Inc.)**: Email is connected per the about page; triage and drafting are not described.

- not this job: AI Job Search (Mads Lorentzen): /gmail-sync proposes application status updates from Gmail; nothing ranks threads or drafts replies, and there is no channel to approve in.
- not this job: asaply (Asaply Inc.): No mailbox connection; the service reads texts from the owner's number only.
- not this job: Asmi (Humint Labs, Inc.): Asmi emails businesses on the owner's behalf; it does not read the owner's inbox.
- not this job: AstrBot (AstrBotDevs): No mail connector and no approval step before any tool call, so a draft cannot be held for a yes.
- not this job: Atticus (An individual developer (GitHub SaarthurR)): Drafts a reply to the email on screen and never sends; no mailbox scan or ranked list.
- not this job: CowAgent (zhayujie): No mode asks before a send; agent_permission_mode defaults to full-access and mail exists only through the agently-cli overlay skill.
- not this job: DeepTutor (HKUDS): No mailbox; Partners answer on IM channels about study material.
- 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: HireAlpha (HireAlpha): Email drafts belong to Coworker, which has not shipped; Friend reads mail only to nag.
- not this job: LearnHarness (No legal entity found. Copyright Amos Blomqvist): No mailbox.
- not this job: Leon (No legal entity found. Copyright Louis Grenard): No mailbox connector and no approval step before a tool call.
- not this job: nanobot (HKUDS): No tool reads the owner's mailbox; the Email channel is a bot inbox that replies to allowed senders over SMTP. No approval step exists in docs or source, so a draft-only run cannot be enforced.
- not this job: PicoClaw (Sipeed): No mail tool, no approval step, and the hooks guide says waiting for a human reply is not supported.
- not this job: szn (Nomadic Futures, Inc.): Does not read or triage the owner's inbox; email is a channel for requests and a way to reach businesses.
- not this job: Tab (Terrasoft, Inc.): Sends approved emails but does not triage the inbox.
- not this job: Talorys (Roc Yu (rociiu) and contributors): No mailbox connection in the code.
- not this job: Wajo (Fo) (Wajo, Inc.): No mailbox reading or triage is documented for Fo; email is a channel and a send capability.

## What stops agents finishing

- approval is a sentence in a SKILL.md or system prompt rather than a gate in code, so a drifted or compacted context sends or deletes anyway
- no way to stop a run from the chat channel once it starts
- mail provider not supported (Microsoft 365 on Duet, iCloud on Poke)

## Around it

Before: personal-assistants.morning-brief. After: email.send, email.unsubscribe.

## Where this comes from

Weight 18 of 100 in Personal assistants: 9 of 40 accounts concern reading, drafting or acting on the owner's mail (acc-02, acc-08, acc-12, acc-13, acc-17, acc-21, acc-22, acc-27, acc-36); it is also where the two publicised failures happened (acc-08, acc-12).

As data: https://www.anchorterminal.com/api/v1/use-cases/personal-assistants.inbox-triage-draft.json, or the MCP tool get_use_case with id personal-assistants.inbox-triage-draft. 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
