ProductUnderstand
Atlasin development
A model of your code, config and traffic, each fact with its source
A model of your system where every claim names its proof.
Atlas is being built to read your code, configuration, live traffic and decisions and turn them into claims: what runs, what depends on what, and which decision governs it. Each claim names the evidence its contract required. Where two sources disagree, you see both. Where nothing has looked, there is no claim, and Atlas says what would settle it.
In development and not usable yet. Book a call to see where it stands.
Engineers answer questions about their own system from memory, old diagrams and whoever is around. An AI assistant fills the same gaps with guesses that read like facts.
Three kinds of source, each to say what kind of evidence it gives (planned), and one fixed set of rules that decides what that evidence can prove.
How it works
- 01
Where it runs
- Planned: the parts that read your system run in your agent, inside your network, read-only. The part that decides what the evidence proves is built on the platform side.
- 02
What it reads
- Kubernetes, Envoy routes, network policy, cargo manifests and the machine itself; your product plans, proposals and decision records in Decisions. Reading code through compilers and parsers is planned.
- 03
What you approve
- Where the running system contradicts a decision, Atlas shows both sides and you decide which one changes.
Each line says how far it is built today. Live means anyone with an account uses it; preview means it runs in early access or as a pre-release; coming means it is written down and not built yet.
What you get
Claims with their proof
Atlas stores only claims and the evidence for them; everything else is worked out from those, with its proof. A model may propose a claim and never writes one. That checking part is built and caught all 26 bugs we planted to test it. It is not available to use yet.
in developmentWhen it held, and when it was learned
The store keeps each claim with two dates: when it was true and when Atlas learned it. Stale is not false: an old proof is kept with the time it held. The store is built.
in developmentContradictions shown on both sides
When the running system goes against a decision (a recorded decision allows three retries, the code sets five), Atlas is designed to show the decision and what the system does and pick no side.
in developmentRead from what runs, by your agent
Readers for Kubernetes, Envoy routes, network policy, cargo and the machine itself, read-only, inside your network. Built and in review: 4,796 records read from our own platform. Not in an agent release yet.
in developmentThe map in the console
The living map on your system, opened to early-access teams after two exit tests: find one true thing nobody told it, and find a planted contradiction.
coming
How one claim is meant to read once Atlas is served: the claim, the evidence behind it, when it held and when it was learned, and a contradiction shown from both sides.
Design sketch
checkout calls ledger over gRPC
verified · current · 3 pieces of evidence
Evidence
- Envoy routecheckout → ledger:7443, cluster ledger-grpc
- Kubernetes Servicepayments/ledger, port 7443 grpc
- cargocheckout depends on ledger-client 2.4
- Held
- since 2026-09-14
- Learned
- 2026-10-02 09:15, by agent-0
Contradiction, both sides shown
ADR-0017 allows at most 3 retries to the ledger
checkout/src/ledger.rs sets max_retries = 5
Atlas picks no side; a person decides.
Example data
A sketch of the planned map, since Atlas is not served yet. The services, files and times are made up.
Every claim on this page has its evidence: the documents that define it, the docs that say how to use it, and the API it runs on.
Proof and links
Documents
- RFC 0086 Atlas core
- RFC 0087 Atlas templates and grading
- PRD 0001 InOrbit Atlas
- ADR 0002 A claim's state on four axes
- ADR 0003 Stale is not false
Not on the public RFCs host yet, so named here without a link.
API
- RESTcomingNo route yet: Atlas is not served.
- MCPcomingNot a tool yet.
The evidence rule: a model saying it succeeded is never evidence. What counts is what was measured, by what, and when.
Your data, your call
The agent reads your system inside your network and read-only; customer code is processed inside your network. A model proposes and never writes a claim.
Logs record who did what and the outcome, never content. The platform's own models are self-hosted; a third-party model provider is one you connect with your own key, and it bills you directly, with no line or markup from InOrbit. Separately, our own DPAs with model providers are in progress, none signed yet. InOrbit is not certified: no auditor has attested its controls yet, and the known gaps are listed.
Questions, answered plainly
Can I use Atlas today?
Not yet. The checking rules and the store are built and the readers are in review; no service serves the map. Early-access teams see it first, on their own system.
Why no confidence score?
Because a number from a model is not evidence. A claim's state is described in several parts, and each names the proof that moved it.
Does it need access to production?
Read-only, through your own agent inside your network. It sends what it observed about your system, never request bodies or secrets.
Next step
See it on your own system
Book 30 minutes with an engineer. We look at your system together and, where it fits, open early access for your team.