AI agents for Gmail: sort the landscape by access model

Why access model, and not a feature list, is the useful axis

Every vendor in this space lists "AI" somewhere in its feature grid, and feature grids read alike: drafting, summarizing, automation, insights. What a feature grid does not show is how the AI gets to your mail, who configured it, and what happens when it is wrong. Two products can both say "AI agent" while one runs entirely inside a vendor's own platform under a plan tier, and the other is an address of its own sending mail with nobody's inbox behind it.

Access model also has real operational consequences, not just architectural ones. Google's own documentation on Gmail delegation states: "Automated processes that frequently access Gmail might also reduce the number of delegates who can access an account at the same time. These processes include APIs or browser extensions that access Gmail frequently." (support.google.com, checked 2026-08-11). How an agent connects is not a footnote — it changes what else can use the mailbox at the same time.

This page is narrow on purpose: it sorts only the AI-agent access models. For the wider category of team inbox software — native Google Groups, Gmail-layer tools, helpdesks, collaboration suites, and governed agent access — see team inbox software: a fit framework.

When these facts were checked

Reviewed on 2026-08-11 — re-verification due by 2026-11-11.

Every vendor fact below was read from that vendor's own published page on 2026-08-11 and is attributed where it is used. Plan tiering and AI feature placement in this category change without notice, so treat anything on this page read after the re-verification date as unverified and check the vendor's own page.

Four shapes an AI agent can take against Gmail

This is not a ranking. Each shape suits a different buyer, and every vendor quoted here publishes real capability. Read what each one states about itself and compare against what your workflow actually needs.

Shape 1 — vendor-bundled: the agent lives inside the vendor's own platform

The most common shape today. The AI agent is a feature of the plan you buy, configured inside that vendor's interface, running under that vendor's own rules.

Hiver lists AI Agents and AI Copilot under its Growth plan and above ("Everything in Free, plus: AI Agents — Resolve requests end to end, analyze conversations, and automate customer operations."), with AI QA and AI Insights added at Elite. Its own positioning: "Your customer operations, run by AI Agents. Governed by you." Hiver also publishes a guardrail and review story: "Guardrails — Decide when AI steps in, when it should hand off and set the bar for what needs human review." and "Admins can review, edit, or disable any AI-triggered workflows." (hiverhq.com/pricing and hiverhq.com/features/ai-agents, checked 2026-08-11). MCP is not described on either page checked.

Front sells its AI as seat add-ons: "Copilot ... $20 / seat / month as an add-on, or included in Enterprise" and "Autopilot ... Starting at $0.05 /conversation." Front also publishes review-before-send language for Copilot: "Review every AI tweak in-line before you hit send, so nothing slips past you." and "Agents review and approve every message before sending, maintaining full control over tone, accuracy, and brand voice." Autopilot publishes the same posture for automated workflows: "Test every Playbook before it goes live, with built-in safeguards and human handoffs when it matters." (front.com/pricing, front.com/product/copilot-ai-assistant, front.com/product/autopilot-customer-service-automation, checked 2026-08-11). MCP is not described on any of the three pages checked.

Gmelius tiers its AI assistants by plan: the Meli plan includes "Meli, your AI Executive Assistant" alongside named assistants for reply composition, sorting and meetings; Growth adds "AI Automation Architect" and "AI Sentiment Analysis"; Pro adds "AI Dispatching Assistants"; Enterprise adds "Bespoke AI Assistants." (gmelius.com/pricing, checked 2026-08-11). MCP is not described on the page checked. Gmelius does not publish where its rule evaluation executes — in the Gmail client or server-side — on the page checked, so this page makes no claim about that.

If this shape fits, DuoInbox is not your answer. A team that wants the agent to arrive with the plan, configured through a vendor's own admin console and supported by that vendor, should buy from this group. All three publish a price today; DuoInbox does not.

Shape 2 — the vendor's AI as an MCP client into your other systems

Missive publishes its AI as working from inside the inbox and reaching out through MCP: "Missive's AI works inside your inbox—reading threads, checking calendars, and drafting replies with full context," and "Ask the assistant to look up a customer's billing history in Stripe, create a Linear issue from a bug report, or pull context from a Notion doc." Setup is customer-configured: "Go to Settings > Integrations > Custom MCP, enter your server URL, and choose an authentication method (none, bearer token, or OAuth 2.0)." Model providers: "OpenAI, Anthropic, and Google supported." (missiveapp.com/ai-assistant, checked 2026-08-11). Missive's own plan matrix lists an "MCP (Stripe, Notion, Linear, etc.)" row, but which plan tier includes it could not be read unambiguously from the served page, so that detail is not stated here. Whether a Missive reply requires human review before sending is also not stated on the page checked.

Shape 3 — the inbox exposed as an MCP server to your own AI client

The mirror image of Shape 2: instead of the vendor's AI reaching out, the inbox itself becomes a tool your own AI client can connect to.

Drag publishes this directly: "MCP Server — Connect Claude, ChatGPT, Gemini, Copilot, Cursor, or any MCP-compatible AI tool to your inbox," and "Works in Claude ChatGPT Gemini Copilot Cursor Claude Code + any MCP client." MCP access is listed under its Plus plan; "API and MCP access" under Pro. Drag also publishes an agent product, currently labeled "Agents Early Access" and "Intelligence Early Access": "Drag Agent — Reads the ticket, pulls context from your tools, takes action. Escalates when it should." Drag meters programmatic use with credits — "Credits apply only to programmatic operations: API, MCP, CLI, and Agent," with AI features in its Chrome extension and web app included in the seat — at published costs of Read 1, Write 2, Send 5, and an AI action 15 credits, against allowances of "2,500 credits/month (free during early access)" on Plus and "10,000 credits/month (free during early access)" on Pro. Drag's own caveat: "Prices are per user, billed annually, and may change." (dragapp.com/pricing, checked 2026-08-11).

Shape 4 — the agent gets its own mailbox entirely

AgentMail is not a Gmail feature at all — it is infrastructure for an agent that is the correspondent, not a worker inside someone else's inbox: "Email for AI Agents." "The email inbox API for AI agents. Create inboxes programmatically, send and receive emails, and manage communications — all via API." Its published plans price by inbox and message volume rather than by seat: Free at $0/month for 3 inboxes and 3,000 emails/month, Developer at $20/month, and Startup at $200/month. (agentmail.to and agentmail.to/pricing, checked 2026-08-11). If your agent is the counterparty rather than a worker inside a team inbox, read own mailbox or shared inbox for an AI agent before evaluating any of the four shapes above.

A fifth shape: a customer-owned agent with governed access, not vendor-bundled or its own mailbox

Where DuoInbox fits. DuoInbox does not bundle an AI agent into its own platform, and it does not give an agent its own mailbox. It gives Hermes, a customer-owned agent, governed mail access to your existing Gmail shared inbox through MCP rather than direct Gmail credentials. Access is scoped by trigger rules that decide when a run starts, one of three permission levels (summarize only, prepare a proposed reply, or send after human approval), a review queue, recipient restrictions, sending limits, an audit log, and an immediate kill switch. This is a fit for teams who already own a shared inbox and want an agent they configure and can stop, not a verdict on the four shapes above. DuoInbox's pricing is not yet published.

If instead your agent needs to send and receive on its own address, independent of any team inbox, DuoInbox is not the answer to that problem — see own mailbox or shared inbox for an AI agent for that decision, and look at dedicated email infrastructure such as AgentMail's published API for that case.

Questions to ask whichever shape you pick

Whichever of these five shapes fits your team, the underlying questions are the same: what credentials does the agent actually hold, what can it reach, and what stops before something goes wrong. Work through the full checklist at AI email agent security questions before committing to a vendor, and see controls for the vocabulary DuoInbox uses for trigger rules, permission levels, and limits. The wider framework behind permissions, review states and audit trails is in AI email agents: a practical governance guide; the per-vendor reads against the same five criteria are on the comparisons hub.

Frequently asked questions

Is an 'AI agent for Gmail' the same thing across every vendor?

No. The phrase covers at least four different access models: an agent bundled into a vendor's own platform and configured within it, a vendor's AI acting as an MCP client that reaches out to your other tools, your own AI acting as an MCP client that reaches into the mailbox through a vendor's MCP server, and an agent with its own mailbox and its own address. Each carries a different answer to what the agent can reach and who configured it.

Does having 'MCP' listed on a pricing page tell you the direction it runs?

Not by itself. MCP (Model Context Protocol) is a connection standard, and it can run in either direction: Missive publishes its own AI as an MCP client reaching a customer-supplied server (Stripe, Notion, Linear), while Drag publishes the inbox itself as an MCP server other AI tools connect into. Read the vendor's own wording for which side of the connection is theirs.

Which of these vendors publish a human-review step before an AI reply is sent?

Front and Hiver both publish one. Front: "Agents review and approve every message before sending, maintaining full control over tone, accuracy, and brand voice." Hiver: "Admins can review, edit, or disable any AI-triggered workflows." Missive's own AI assistant page does not state whether a reply requires review before sending, so that page records it as not stated rather than assuming either way.

Where does DuoInbox fit among these four shapes?

DuoInbox is a fifth shape: a customer-owned agent (Hermes) given governed access to your Gmail shared inbox through MCP, rather than a vendor-bundled agent or direct Gmail credentials. It fits a team that already owns a shared inbox and wants to configure trigger rules, permission levels, and a review queue around an agent it owns. DuoInbox's pricing is not yet published.

Front, Hiver, Gmelius, Missive, Drag, AgentMail, Gmail, and Google Workspace are trademarks of their respective owners, used here descriptively only. DuoInbox is an independent product and is not affiliated with, endorsed by, or sponsored by any of them. Every quoted fact above was read from that vendor's own current public page on 2026-08-11 and may have changed since; check the vendor's page directly before relying on a figure or a plan-tier placement.