Gmail sorts by time.
The "Primary" tab knows nothing about who you know. Your last forty threads with Alice sit next to a calendar reminder and four newsletters.
An essay in four parts, on email.
Your inbox
is already
a deal-graph.
You already have a deal-graph — it's buried in three years of signatures, intros, and threads. Twang extracts it. Ask "where does Alice work?" or "who introduced me to Anthropic?" — answer in 200ms. No CRM. No data entry. Just your inbox, finally legible.
Private alpha · macOS 13+ · Free during alpha
Inbox · 14
Jenna Liu
Term sheet v3 — redline
Paul (Portfolio)
Updated cap table
Sarah Chen
Intro — Rivera (YC W26)
Founder digest
Weekly roundup
I.
The problem
You run 300 relationships through your inbox. The inbox doesn't know any of them.
Every relationship you have is documented in your inbox — every signature, every warm intro, every CC. None of it is structured. Gmail sorts by time, not by who you know. Affinity is a separate CRM you forget to update. Apple Mail and Superhuman treat signatures as visual noise.
A CRM is a database you don't keep.
Affinity, Attio, Salesloft — all built around manual entry. By the time you update the contact, the deal is closed and the data is wrong.
AI inboxes treat signatures as noise.
Superhuman makes your email faster. Shortwave replies for you. Neither one reads who wrote the email, where they work now, or who introduced you.
Twang is a different shape.
II.
The signals
Every email contains relationship structure.
Every signature is a "works at" assertion. Every warm intro is an "introduced by" edge. Every CC ties two people together. Every calendar invite confirms a working relationship. Twang reads all four signals at sync time and builds your deal-graph.
i.
Signature
Parsed at sync time. "Alice Liu, Sr. PM, Anthropic" → works_at edge, confidence 1.0.
ii.
Intro
Detected via nineteen phrasings — "meet X", "looping in", "intro:". introduced_by edges with provenance.
iii.
CC
Every CC pairs two organizations. Confidence scales with frequency and recency.
iv.
Meeting
Calendar attendance bumps works_at confidence. Working relationships beat domain inference.
Four signals, scored 0.5 to 1.0 by source. The graph compounds with every email.
III.
Feature 01 · Speed
Every action is one keystroke away.
Twang's keyboard model matches what your hands already know. JK to move, E to archive, R to reply, ⌘K for anything else.
Navigate
JKG IG SEsc
Compose
CR⌘↵⇧⌘↵
Triage
E#HM
Search & jump
⌘K/@G FG D
Every shortcut you already know works here. Twang adds lane jumps and search operators on top.
IV.
Feature 02 · The query
Ask your inbox anything about your relationships.
Open ⌘K. Ask "who works at Anthropic?" — 200ms to a list of contacts. Ask "where does Alice work?" — 200ms to her current company. Ask "who introduced me to Acme?" — 200ms to the original intro thread. Pure SQL retrieval. Zero LLM hallucination. Citations you can click.
JL
Jenna Liu
General Partner · Index Ventures
87Relationship
3This week
4dSince reply
Term sheet v3 — redline9:42 AM
Re: Rivera intro thanksApr 17
Partners meeting notesApr 14
Forward — "who works at X?"
List every contact in your inbox at any company. Sorted by confidence and recency. Each row links to the email it came from.
Reverse — "where does Alice work?"
Surface all current and past works_at edges for a person. Job changes show up as new edges; the old ones don't disappear.
Intro — "who introduced me to Acme?"
Find the warm-intro thread that started a relationship. Surfaces the sender, the date, and the original email.
V.
Feature 03 · Drafts
We draft. You decide.
Twang writes a reply grounded in your sent mail — pulling the three closest "received → your reply" pairs from your past as exemplars. Open the draft. Edit if needed. Press send.
Draft · for Jenna Liu 2:04 PM
Jenna —
12x works if we can hold 1x non-participating. Free 3–4pm tomorrow for a jump, or I can send the redline back tonight if that's faster.
Navot
Why this draft? Jenna mentioned 10x last thread, and your last reply on term sheets held the line on non-participating prefs.
- Grounded in your sent mail. Retrieves your 3 closest reply exemplars via FTS5 search.
- Transparent. "Why this draft?" on every output.
- Never sent without you. Messenger, not author.
VI.
Feature 04 · Privacy
Your inbox doesn't have to leave your Mac.
Most users run Twang in cloud mode with hosted models. If you need stricter privacy, Privacy mode keeps all inference on your Mac. The cloud never sees your email.
Cloud mode · default
Your Mac → OpenRouter → Your Mac
Mistral Small 3.1 for classification, Llama 3.3 70B for drafting. About $1.50/user/month of inference cost. Fastest and highest quality.
Privacy mode · opt-in
Your Mac → Your Mac
Gemma 4B on your Apple Silicon GPU. All inference on-device. Lower throughput. Email never leaves your machine.
- Encrypted local database (SQLCipher).
- OAuth tokens in macOS Keychain.
- Zero telemetry. You opt in to every feature flag.
- Switch modes per session.
VIII.
Security
Built for a bar higher than email.
Local database.
Your emails, contacts, and drafts sit on your Mac in an encrypted SQLCipher database. Twang syncs them from Gmail. It doesn't mirror them to our cloud.
macOS Keychain.
OAuth tokens and API keys live in the Keychain. We don't see them. You can revoke Twang in Google's security panel at any time.
You pick the boundary.
Cloud mode sends thread content to OpenRouter for drafting and triage. Privacy mode keeps inference local. You pick.
IX.
FAQ
Questions people actually ask.
What's a deal-graph?
Every email you receive carries a small piece of relationship structure — a signature line with someone's company, a warm intro pairing two people, a CC that ties two organizations together. Twang reads all of those signals across your full inbox and assembles them into a queryable graph: who works where, who introduced whom, who knows whom. Then you can ask "where does Alice work?" or "who introduced me to Anthropic?" and get the answer from emails you already have, without maintaining a separate CRM.
Does Twang work with Gmail?
Yes. Twang is a Gmail client first. It reads and writes through Gmail's API. Google Workspace accounts are fully supported. Outlook is on the roadmap for 2026.
Does Twang send email on my behalf?
No. Twang drafts every reply, but you press send. We don't believe AI should speak in your voice without your review.
Where does my email live?
In an encrypted SQLCipher database on your Mac. We don't mirror your inbox to our servers. See security for the full data map.
What does AI mode mean?
By default, Twang sends thread content to hosted LLMs (Mistral Small 3.1 for triage, Llama 3.3 70B for drafting, both on OpenRouter). Switch to Privacy mode to route all inference to a local model on your Apple Silicon GPU.
How much will Twang cost?
Free during alpha. When we open public access, expect a per-seat model priced like premium email clients. We'll announce pricing at least 30 days before we start charging alpha users.
What's the current status?
Private alpha, invite-only. macOS 13+ on Apple Silicon or Intel. Windows and mobile are not currently on the roadmap; Outlook support is targeted for later in 2026.
How is Twang different from other AI email tools?
Most AI email tools try to be autonomous. They reply on your behalf. Twang doesn't. You press send on every reply.
Who's behind Twang?
Twang is built by Navot Volk (Partner at GroundUp Ventures) and a small team. We built the inbox we needed for our own work.
X.
Ready?
Ready for your inbox to do the work?
Request alpha access. We review every request personally. If it's a fit, we'll send an invite and book a 20-minute setup call.
Request Alpha AccessFree during alpha · macOS 13+ · Or write to navot@groundup.vc