Upstream

Upstream is an AI-native email client designed around a stronger premise than ordinary AI-assisted drafting: agents should be first-class participants in the inbox. The product triages incoming mail, separates low-priority noise, prepares replies in the user's writing style, tracks unanswered threads and can carry out inbox-related tasks such as finding receipts or scheduling meetings.¹²

The current product sits on top of Gmail rather than replacing email as a protocol. Y Combinator's launch description says Upstream rebuilt the email client as a surface agents can read, write and act on, with inbox organization, custom voice prompts, built-in task handling and smart follow-ups.¹ The iOS app independently confirms the main consumer-facing workflow: automated triage, drafts tuned to the user's voice and follow-up generation, with a Noise split for low-priority mail and support for multiple accounts added during 2026.²

The agent architecture is the most technically differentiated part of the product. Upstream exposes an MCP server with 44 tools that let compatible assistants browse inbox splits, search and summarize threads, compose and reply, star, snooze, label, archive and create rules.³ The server is available over Streamable HTTP and can be used from MCP-compatible clients, including via a bridge for clients that expect stdio.³

Authentication uses personal access tokens created inside Upstream. A token can access the user's inbox, threads, messages, labels, channels, rules, settings, contacts and organization members and can send replies or compose mail on the user's behalf.⁴ Upstream recommends one token per client so access can be revoked separately. The MCP server acts as a proxy to the account rather than storing a separate copy of the email data.⁴

The broader context layer is designed to reduce generic drafting. Upstream's YC launch material describes connections to tools such as Granola, Notion, Google Drive and Calendar so agents can use work context before drafting or acting.¹ System prompts are editable, allowing teams to tune behaviors such as what counts as needing a reply or how drafting should work.

Team collaboration is also part of the positioning. Rather than forwarding messages or discussing them in a separate chat product, Upstream supports channels and shared context around email workflows. Chrome Web Store material emphasizes channels grouped by project, client or topic alongside AI drafts and prioritization.⁵

The product remains narrower than a universal email replacement. Current independent launch coverage says Gmail is supported while Outlook compatibility is still planned, and Android support is also still in development.⁶ Upstream is available on web, desktop and iOS, but the provider/platform gap matters for organizations with mixed Microsoft and Google environments.

Upstream has a meaningful free entry point. Y Combinator's launch post says setup is free and notes that beta users had replaced Superhuman with Upstream's free plan.¹ Product Hunt currently lists Upstream as Free and promoted a launch offer for one free month of Pro.⁷

A current secondary review published in August 2026 lists Core as free for teams of up to five people, Pro at $20 per user per month when billed annually and Max at $60 per user per month.⁸ Other current startup-market material corroborates a freemium model with per-seat paid subscriptions, while Tech.eu confirms the existence of both a free tier and a paid Pro plan.⁶

The exact live pricing UI was not independently readable in this cycle, so these amounts should be treated as current secondary commercial evidence rather than immutable first-party facts. Upstream is moving quickly, and buyers should confirm the current monthly/annual billing terms and AI-usage limits before purchase.

The free tier is strategically important because the product competes with established email clients where switching cost is behavioral as much as financial. A user can test whether Upstream's triage, drafts and follow-ups materially reduce inbox time before paying for expanded AI or team functionality.

Value depends less on the headline seat price than on how much autonomous work is genuinely useful. If users routinely rewrite the drafts, mistrust triage or disable follow-ups, the product becomes a more conventional inbox with AI decoration. If the agents reliably clear low-value work and surface the right threads, even a premium per-seat price can be inexpensive relative to professional time.

Upstream has more independent evidence than many products of similar age, but the samples remain small. Apple's U.S. App Store currently shows 3.7/5 from three ratings. One review calls the product a strong alternative to Shortwave and Superhuman, while another argues that the AI layer is oversold and characterizes it mainly as a Gmail organizer with reply writing.² Three ratings are far too few for a stable numeric product score, but the disagreement is useful qualitative evidence about the product's core value proposition.

The Chrome Web Store is directionally more positive: the Upstream extension currently shows 5.0/5 from six ratings and several hundred users.⁵ Again, the sample is too small for a defensible ecosystem-wide numeric rating, and extension ratings measure only one access surface.

Product Hunt shows 5.0/5 from one review and roughly 1.8K followers, alongside strong June 2026 launch rankings.⁷ The launch traction supports market interest but not long-term reliability. Product Hunt is also launch-adjacent and maker-heavy, so the single rating is retained qualitatively rather than treated as a mature user-review signal.

Tech.eu provides unusually useful independent evidence because its reporter says she tried Upstream while handling a high-volume inbox.⁶ The review describes familiar Gmail-like interaction, AI prioritization, drafts, follow-ups, channels and editable behavior. It also highlights an important control: AI-generated drafts require user approval before sending in the normal product workflow.⁶ That is stronger evidence than a feature directory, although it remains one journalist's short launch-period test.

Privacy and access control deserve close attention because email is unusually sensitive data. Apple's disclosure says Upstream may collect and link contact information, contacts, user content including emails/messages and media, identifiers, usage data and diagnostics to the user's identity.² The Chrome Web Store similarly states that the extension handles personally identifiable information and declares that data is not sold or used for unrelated purposes.⁵

Upstream's public Privacy Policy describes standard collection/use/sharing and user-rights language, but the most operationally important security facts for agent access are in the MCP authentication documentation.⁹⁴ A personal access token can expose the same inbox data a user sees and authorize composing, replying and trashing or organizing messages. Upstream recommends keeping tokens private, using one token per client and revoking compromised tokens immediately.⁴

Tech.eu reports the founder's statement that Upstream does not train AI models on customer data; message history may be analyzed temporarily for contextual drafting but is not turned into a training dataset.⁶ This is meaningful, but it is still a company statement reported through an interview rather than an independently audited assurance report.

The risk profile therefore comes from capability as much as privacy policy. A read-only summarizer can leak information; an agent-connected inbox can additionally send, archive or delete. Organizations should pair Upstream's token separation and confirmation flows with clear rules on which external MCP clients can connect, who owns tokens, what actions require approval and how access is revoked when a device or agent is retired.

Upstream is best suited to professionals whose inbox is itself a work queue. Founders, executives, sales leaders, recruiters, investors and operators who receive enough mail that triage and follow-up consume meaningful daily attention are the clearest audience.¹⁶

Gmail-centric small teams can gain additional value from shared channels and contextual collaboration without moving discussions into a separate messaging product.⁵ The free tier also lowers the barrier to testing the workflow with a small group before broader adoption.

Users already working heavily with Claude, Codex or other MCP-capable agents are another strong fit. Upstream gives those agents a structured email interface rather than forcing browser automation or ad-hoc Gmail integrations.³ The 44-tool surface is broad enough for real workflows, not just search.

Upstream is less suitable for organizations standardized on Outlook/Microsoft 365 today, Android-first users who need full native parity, or enterprises that require mature independent assurance and a long operating history before allowing agent tokens into email.⁶

It may also be a weaker fit for users whose primary need is ultra-fast manual keyboard triage rather than automation. Traditional premium clients such as Superhuman optimize the human operator; Upstream's strongest differentiation appears when the user is willing to delegate organization, drafting and follow-up to agents.