OXYGENOXYGEN

For AI startups

GTM infrastructure for AI startups

Everyone in your category now sends AI written email, so writing is no longer the advantage. Research depth is, and that is a workspace problem.

try
Start freeRead the docs

Selling AI to people building AI is a strange market. Your buyer runs the same models you do, spots a generated paragraph immediately, and has already deleted four emails today that opened with their own funding announcement.

What still works is doing the reading. A research column can visit a company's site, its docs, and its posts, then answer one specific question about fit. The answer lives on the row with its sources, so a person can check it before anything sends.

Who this is for
Seed to Series A AI companies selling software rather than services: agents, copilots, model tooling, or applied AI for one business function. Four to fifty people.
What it looks like today
Every competitor generates the same personalized paragraph from the same profile blurb. The cost of a mediocre email has fallen to zero, and what those emails produce has fallen with it.
What changes
Research becomes a column instead of a prompt. Each account is read from the open web, graded against your rubric, and the answer sits beside its sources, so you can see which rows deserve a human sentence.

Example searches

Ask for it the way you'd say it

Every search below runs on the same hosted Tables, with the cost previewed before anything is spent.

  • Funded and building

    list companies that raised in the last six months and are hiring machine learning engineers

    Round dates and postings sit on the same account, so recency and intent can be sorted against each other.

  • Fit, judged by a rubric

    read each company site and score how much their product overlaps with what we replace

    The score arrives with the pages it read, so a low grade can be argued with rather than taken on faith.

  • The technical buyer

    find the cto or founding engineer at each of these companies with a verified email

    At this size the technical founder is often the buyer, and the table records which role it resolved.

Plays

Three motions you can run this week

Each one is a chain of Oxygen primitives — the same hosted objects your workspace already has, composed.

  • Research at list scale

    A graded list where every row was actually read, with its sources attached.

    1. Tables
    2. Knowledge Graph
    3. Records
    1. 1Write the rubric once into the workspace wiki so every run scores the same way.
    2. 2Run a web grounded AI column over the table, then promote the rows that pass.
    oxygen tables create

  • An agent that keeps the list current

    A bounded worker that refreshes research across your target accounts.

    1. Agents
    2. Tables
    3. Observability
    1. 1Give the agent a goal, a capability scope, and an approved credit ceiling.
    2. 2Read the run log to see every tool call, cost, and output it produced.
    oxygen agent run

  • One sentence a human wrote

    Generated research, human opening lines, and a cadence that stops on reply.

    1. Tables
    2. Sequences
    3. Messages
    1. 1Keep the research in columns and write the first line yourself on rows that matter.
    2. 2Enroll, then triage answers in the shared inbox rather than a personal mailbox.
    oxygen sequences start

Capabilities

What you get

  • Research as a column, not a prompt

    One question, asked of every row, answered from pages the run actually read, with those sources kept beside the answer for anyone to check.

  • Agents with a governor

    A workspace agent runs under an approved credit ceiling and a scoped capability set, and every tool call it makes is recorded as a durable run.

  • Grounded in your own context

    Positioning, ICP, and rubrics live in the workspace wiki, so drafts and scores read one source instead of whatever happened to be in the prompt.

  • Start free, upgrade to send

    Tables, workflows, drafts, and defining an agent are unmetered. Agent runs, enrichment, and AI columns draw credits under a ceiling you approve before anything starts.

Instead of

The stack this replaces

Native, on one contract and one credit balance — not another tab wired to the last one.

  • Clay
  • Apollo
  • Instantly
  • HubSpot
  • n8n
  • Zapier

Data sources

What the data actually comes from

Every value lands with its provider and cost recorded on the cell.

  • PredictLeads

    Funding and hiring events, which is how a company that just decided to spend becomes visible.

  • Crustdata

    Headcount trends and postings per company, useful for separating a real team from a launch page.

  • Exa

    Neural web search for grounding research columns in pages the run has actually read.

  • Blitz API

    Resolves the technical founder or engineering lead behind the account so research reaches a person.

Run these on Oxygen's managed credits, or connect your own provider keys and pay the vendor directly — the same columns, the same runs, the same provenance either way. See every integration.

Limits

Where this stops

  • A research column is only as good as its question. Ask something vague and you get a confident paragraph nobody can check, so write questions you could verify yourself in a minute.

FAQ

Questions people ask first

Is this just another AI writing tool?
No. Writing is the smallest part of it. Sourcing, research, grading, sending, and replies share one database, so a claim in an email can be traced back to the page it came from.
How do we stop an agent spending money we did not approve?
Tool access and spending are governed separately. An agent run carries an approved ceiling, a scoped capability set, and a durable log, and the platform refuses anything outside them.
Our ICP keeps changing. Does that break the lists?
It does not, because the definition lives in the workspace wiki rather than inside a prompt. Change it once and the next run scores against the new version, while older runs stay readable.
Can the research cite where it got an answer?
Yes, and it should. The pages a run read are recorded with the answer, which is the difference between a score you can defend on a call and one nobody is able to check.
How much does research over a thousand rows cost?
Preview it first. AI columns and web grounding are metered per row, the run states its ceiling before starting, and you can sample twenty rows before committing to the rest.
Do we need to connect our own model keys?
Not to begin with. Managed access is the default so nothing has to be wired up on day one, and any lane that supports your own key can use it once you want that control.

Put your GTM motion on one stack

Sign up, get a working workspace with a one-time credit grant, and run the first play end to end without wiring five tools together.

Start freeSee pricing