Shortwave
Gmail-only AI email client with custom embeddings indexing the whole mailbox, agentic tasklets and MCP tool connections. The permanent free plan was withdrawn in 2026.
Overview
Built by people who came out of Google Inbox, and it shows in the keyboard-driven, thread-centric design. This is a client rather than a mail provider, connecting through the Gmail API for personal Gmail and Google Workspace accounts. Apps for macOS, Windows, iOS, Android and the web all share the same index.
The engineering worth noting is the retrieval layer. Instead of stuffing recent mail into a prompt, Shortwave maintains custom embeddings of the mailbox in a vector index and runs that beside Postgres and Elasticsearch, so the assistant can answer questions spanning years of correspondence and point at the threads it used. On top sit tasklets, agent runs that triage and draft against rules you write in plain language, and MCP client support added in May 2025 that lets the assistant reach outside tools. Model choice is tiered by plan, Claude Sonnet at the entry level and Opus with extended thinking at the top.
Pricing moved against users in 2026. The permanent free plan was withdrawn quietly around mid-year, leaving a 14-day trial and three paid tiers. Business is $30 a seat monthly or $24 annually, Premier $45 or $36, Max $120 or $100. Those tiers gate daily AI request quotas, search history depth and how many threads a single query may consider. It remains Gmail-only, which rules it out for most Microsoft shops regardless of how good the retrieval is.
Key Features
- ✓ Retrieval over the entire mailbox using custom embeddings in a vector index, run alongside Postgres and Elasticsearch
- ✓ Answers cite the actual threads they drew on rather than paraphrasing the last few messages
- ✓ MCP client support added in May 2025, so the assistant can call external tools and data sources from the inbox
- ✓ Tasklets, multi-step agent runs that triage, draft and file mail against rules written in plain language
- ✓ Model tiering by plan, Claude Sonnet on the entry tier and Opus with extended thinking at the top
- ✓ Native apps for macOS, Windows, iOS and Android plus web, all on the Gmail API
Where it holds
- • Search genuinely works across years of mail, which is the one thing Gmail's own AI still handles badly
- • The retrieval design is public enough to reason about, unlike most inbox AI that just calls it magic
- • MCP support means it plugs into tooling you already run instead of a fixed integrations list
- • Keyboard-first design from ex-Google Inbox people, fast in a way most AI email clients are not
Where it breaks
- • Gmail and Google Workspace only. No Outlook, no IMAP, no Exchange, and Outlook has been in testing since January 2026 without shipping
- • The permanent free tier disappeared mid-2026 with no announcement, so the entry price is now $24 a seat annually
- • Daily AI request quotas are vague, roughly 150 to 300 on Business, and heavy users run into them
- • Indexing your full archive means handing years of correspondence to a third party, a hard sell in a regulated org
My Take
Email clients are a graveyard of good intentions, so it is worth naming the actual thesis here: index the whole mailbox with custom embeddings, then let a model answer over that index instead of over the last twenty messages. It holds up on exactly the query that defeats Gmail search, the vague one about a thread from two years ago whose subject line nobody remembers. Two caveats before committing a team. It is Gmail-only, with Outlook in testing since January 2026 and still unshipped, and the permanent free plan vanished mid-year with no announcement, so the floor is now $24 a seat billed annually.
Quick Info
- Pricing:
- paid
- Openness:
- Proprietary
- Starting at:
- No permanent free plan since roughly mid-2026, only a 14-day trial. Business is $30 per seat monthly or $24 billed annually, Premier $45 or $36, Max $120 or $100. Tiers gate the daily AI request quota, search history depth and how many threads one query may consider: 50 on Business, 100 on Premier, 150 on Max. Business also caps AI-powered filters at 3 against 50 on Max.
- Added:
- Aug 2026
- Updated:
- Aug 2026
Use Cases
Judge it on your own work
The notes above say where Shortwave holds and where it breaks. The fastest check is your own workload.
Visit website ↗