OXYGENOXYGEN

Workspace knowledge

The workspace wiki every AI action reads first

Stateless assistants forget your positioning every session. This is where the company profile, ICP, offers, voice, and rubrics live, so every AI step in the workspace reads them by default.

try
Start freeRead the docs

Every AI tool you use starts from nothing. You paste the same positioning into a prompt, get something almost right, tweak it, and paste it again next week. The missing piece is not a smarter model. It is durable company knowledge that lives in the workspace and gets read without anyone remembering to attach it.

Two shapes share one store. A typed core that AI actions consume programmatically: company profile, ICP, personas, offers, positioning, proof points, disqualifiers, voice, and scoring rubrics. And wikilinked pages for research, meeting notes, and campaign learnings. Beneath both sit immutable sources and a log naming every revision's author and run.

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.

  • Write it down once

    record our positioning, our three offers, and who we disqualify, then use it everywhere

    Typed sections inject into AI actions automatically, and pages join by retrieval when the topic matches.

  • What did the campaign teach us

    summarize what replies told us about objections in the last campaign

    Outcome auto-filing is live on dev and switched off in production today, so file the replies worth keeping as sources yourself and let synthesis distill them into a cited page.

  • Who changed this and why

    show me every revision of the voice page and what each one was based on

    Revisions carry actor, run, timestamp, and cited sources, so a wording change is reviewable rather than mysterious.

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.

  • Seed the wiki, then fill the stubs

    A new workspace starts with a real page skeleton instead of an empty text box nobody opens.

    1. Knowledge Graph
    2. Records
    1. 1Seed the starter set: a start-here guide, a conventions page, and draft stub hubs.
    2. 2Fill the ICP, offers, and voice stubs first, since those three ground almost everything.
    oxygen knowledge seed

  • Close the learning loop

    What replies and closed deals taught you becomes a page the next campaign actually reads.

    1. Messages
    2. Knowledge Graph
    3. Sequences
    1. 1File the reply excerpts and won-lost outcomes that taught you something as immutable sources on the campaign page.
    2. 2Synthesize them into a campaign-learnings page the next brief is written from.
    oxygen knowledge synthesize

  • Gate the copy customers read

    Working notes update freely while brand and positioning changes wait for a person.

    1. Knowledge Graph
    2. Publishing
    1. 1Let ordinary pages write through as logged, revertible revisions.
    2. 2Review canonical voice and positioning edits in the proposal queue before they land.
    oxygen knowledge proposals

Capabilities

What you get

  • Grounded by default

    The typed company profile and the canonical voice and positioning pins are injected into AI columns, sequence drafts, and agent runs without anyone attaching them. Free-form page retrieval runs on dev and is held back in production until a human signs off on how well it works.

  • Provenance on the output

    Every grounded result records which revisions grounded it. When a draft says something odd, you can trace the sentence back to the page that taught it.

  • A risk-tiered write gate

    Working pages auto-write as logged revisions anyone can revert. Canonical voice, brand, positioning, and pinned playbooks wait in a proposal queue for a human decision.

  • A local mirror

    Sync the wiki down to files, edit it in the editor you already use, and push it back. The workspace stays the source of truth and your terminal stays the interface.

Boundaries

What Knowledge Graph does not own

The knowledge graph describes kinds of customers. Actual customers are Records. Working datasets are Tables, message bodies are Messages, and an installable executable motion is a Recipe rather than a page you write.

Data sources

What the data actually comes from

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

  • Exa

    Search used when a page needs a fact from outside the workspace, recorded as a cited source on that page.

  • Firecrawl

    Reads your own site and the pages you point at, which is how a first company profile gets filled at all.

  • Linkup

    A second research path for pages where one engine's view of the web is not enough to write from.

  • Pa

    Parallel

    Deeper research for a page worth an hour of somebody's reading rather than a single snippet.

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

  • Grounding is only as good as what you wrote. An empty wiki produces generic copy very politely, and the first hour spent on ICP, offers, and voice is the highest-leverage hour in the workspace.
  • Synthesis is one low-cost model call over evidence you already have, not a research project. It distills what your own replies and outcomes said; it does not go and find new market facts for you.
  • Automatic grounding is narrower in production than on dev: the typed profile and canonical pins inject, while free-form page retrieval and outcome auto-filing stay switched off pending a human efficacy review.

FAQ

Questions people ask first

Why not keep all this in a docs tool?
Because nothing reads it. A page here is injected or retrieved by the AI steps that write your emails, score your rows, and run your agents, and the revision that grounded an output is recorded on the output itself.
What gets injected automatically?
The six-section company profile plus the canonical voice, brand, and positioning pins, injected without anyone attaching them. Per-section ICP and persona objects an action could resolve one at a time are still in build. Free-form pages join by retrieval on topic match, which runs on dev and is off in production for now.
Does an agent need permission to write here?
It depends what it writes. Observations file as immutable sources without approval, because they are facts about what happened. Canonical customer-facing copy and typed-core synthesis always wait for a human in the proposal queue.
Can I edit this from my own editor?
Yes. A local mirror syncs pages down as files and pushes changes back, so you can work in the editor you already live in. The hosted workspace stays the source of truth for what actually grounds a run.
What counts as a source?
An immutable raw observation: a reply excerpt, a meeting note, a won or lost outcome, a page a researcher read. Sources are cited by synthesized pages, which is what makes a claim on a page checkable later.
Does writing a page cost credits?
Writing, reading, linking, and searching are unmetered. Credits are drawn by the model calls behind synthesis and by any research provider a page asks for, and those show up as ordinary runs.

Run the workspace wiki every ai action reads first on your own workspace

Sign up, get a working workspace with a one-time credit grant, and run the first play from the web app, the CLI, or your AI assistant over MCP.

Start freeSee pricing