Gatlas gives the people and agents running your companies better context.
Your companies’ knowledge is spread across the tools each one runs. Gatlas connects those tools and describes what they hold, so your agents and your team can act on it. Bifrost Partners LLC operates Gatlas, and we build it in New York City and Copenhagen.
An agent knows only what it is handed
A group’s books are correct, company by company, in whichever system each company runs. What made them usable was a person: someone who knew which spreadsheet was current and which figure the accountant had restated. That knowledge was never in a tool.
More of the reading now goes to agents, and an agent knows only what it is handed. Handing it one system of one company keeps getting cheaper, since every layer of the data stack now sells itself as infrastructure for agents and model vendors ship their own connectors to the largest ledgers. None of that reports gross margin across companies that keep different ledgers in different currencies, by the group’s own definition.
Context is data that is safe to act on, because its definitions and its permissions travel with it. Two parts of it stay scarce: a definition that holds across companies, and the boundary that decides who may see each one. Gatlas is built for both, because an agent repeats a wrong definition at scale.
We are building better context, and it has seven properties
- Complete
- Every tool a company runs will be connected. A company library will keep what lives in no tool, such as company memory, shared skills, brand files and the variables everyone should read the same way, written once and versioned.
- Relevant
- Only what serves the work will move forward. A connection will filter at the source, and your workspace, the account your companies sit in, will decide what reaches an agent.
- Consistent
- Each number will have one shared definition: you will write gross margin once, and Gatlas will render it for every company and for the group, across ledgers and currencies. A company that needs its own will get an explicit override, never a silent fork.
- Findable
- Context will be fast to reach and described for a model that has never seen your workspace, and one search will cover your files and your datasets, the curated tables an agent reads, alike.
- Legible
- Artifacts, the dashboards and documents built from your data, will be shaped for a person and follow your brand. Gatlas will keep what you send as an edition, a frozen copy, so a board pack sent in March still says in April what it said in March.
- Delivered
- Activation, everything that puts context to work outside Gatlas, will carry it to wherever the work happens, on a schedule or the moment a number moves. A reverse sync will be one form: a dataset pushed into another tool and kept current there.
- Permitted
- The control layer, meaning settings, members, permissions and billing, will give each reader exactly what their grant allows, down to one branch of your company tree.
Every part of Gatlas has to improve at least one of the seven. A feature that improves none does not belong.
Each position we hold costs us something
A wrong number that looks right is worse than an outage.
Someone notices an outage in minutes. A failed lookup rendered as zero, a truncated page summed as though it were whole, a stale figure with no date beside it, an answer that quietly left out a company: each one gets quoted in a board pack and found, weeks later, by someone who trusted it.
So a missing value renders as missing, never as zero, and a partial or stale result says so. Gatlas will sometimes show you less than a competitor would. That is the choice, taken deliberately.
Defining the rows is the product.
Moving rows out of an accounting system is solved, by us and by everyone else. What is not solved is telling an agent that a particular commission account belongs in sales and marketing rather than against revenue, in this company, from this year forward, and holding that meaning across every company in the group.
A dataset is finished when its description and column notes let an agent use it correctly the first time. That writing is slow work, and it is the work we do first.
No company changes how it runs.
Each company keeps its own tools, and Gatlas reads them where they are. Portfolio monitoring collects the figures a company reports, not its ledger, and a new group-wide system asks every subsidiary to migrate. We do neither.
The cost is dependence on every provider’s API. Writing a connector is cheap; shipping one is not, because a provider can cap unreviewed apps or offer no API at all.
Whoever owns a login connects it.
The person who owns a login authorizes it themselves, through a setup link sent to them, so nobody at the group has to hold a company’s credentials. Once stored, a company’s credential is used only to reach the tool it came from.
This is slower than asking a partner to paste an API key into a form, and we lose time to it in every onboarding. We think it is the correct trade and we are not planning to revisit it.
Depth comes before breadth.
Whatever reads from Gatlas, a person or an agent, inherits the quality of the definitions beneath it. So we deepen curation and search before we add sources and surfaces, and each new connector has to say whether it serves a tool our customers already run or opens a market we cannot reach yet.
Breadth is what gets a company into Gatlas, so putting depth first costs us reach.
A feature that works only in the dashboard is unfinished.
Most people will reach Gatlas through an agent, so the real reader is a model that has never seen your workspace. Everything you can do in the dashboard should also work over MCP, the open standard agents use to connect to tools, and some customers should never need to open the dashboard.
The cost is building every feature for two readers. Your agent can already build datasets and artifacts over MCP; administering members, keys, companies and billing still needs the dashboard.
If you think one of these is wrong, we want the argument, and it is the fastest way to change what we build: team@gatlas.ai.
Jonathan R. FeilbergFounder and CEO, Gatlas. New York.
A small company runs Gatlas
- Bifrost Partners LLC.
- New York City, United States, and Copenhagen, Denmark.
- Our own. The first datasets and guardrails came out of our own month ends, which is also why we find the embarrassing bugs before you do.
- Each agent connection is read-only or read/write, and every call it makes is logged. The Gatlas agent in Slack only reads your data, and an admin decides where it answers.
- Deliberately small. If that is disqualifying for your portfolio, it is better that you find out on this page than in month three.
- No published list of roles. If giving agents better context across a group of companies is your kind of problem, write and say what you would build. team@gatlas.ai.