KAOS — the Kelvin Agentic Operating System
The alternative to Harvey and Legora. Self-hosted. Auditable. Every answer traceable to a page.
from kaos_pdf import parse_pdffrom kaos_content import search_document
# Pull every change-of-control clause from a signed deal.doc = parse_pdf("merger_agreement.pdf")hits = search_document(doc, "change of control", top_k=10).results
for hit in hits: print(f"page {hit.page}, score {hit.score:.2f}") print(f" {hit.text[:120]}…")
# page 14, score 8.21# Upon a Change of Control, each Holder may require the Company to# repurchase its Notes at 101% of the principal amount, plus accrued…How it compares
Six capabilities a partner or CIO actually asks about, the day they evaluate a legal-AI vendor.
What firms do with it
Each example below runs on documents you control, with a citation trail you can hand to a partner, a regulator, or a court.
Ask a legal question, get an answer with the cases, statutes, and pages it relies on — each linked to the exact source so you can check it in one click. When a cite doesn't support the claim, the system flags it instead of hiding it. The final judgment stays with you; this is built for review, not autopilot.
Generate a memo, summary, or first-pass draft from your own documents. Each sentence in the output points back to the page and paragraph it came from, so a partner or compliance reviewer can confirm every line.
Run extraction, Q&A, and summary across thousands of contracts. Set a per-matter spending cap before the work starts. Watch the meter as it runs. Stop the job when the cap is hit, with the partial output preserved.
Why KAOS
The architecture choices behind these answers are written up for engineers and CIOs on /why-kaos.
The system runs wherever you put it — a laptop, your own servers, or your own cloud account. Documents and prompts never leave your environment. Privileged work product stays under your DMS, your retention rules, and your audit logs.
Every answer carries a pointer back to the page and paragraph that produced it. When the system doesn't have the evidence to answer, it refuses and says so. No silent guessing, no invented cases, no fabricated quotes.
The code is public — one repository per package. Your security team can read it, your IT can fork it, and the license is Apache-2.0, so you are not locked into a vendor's release schedule or a vendor's pricing. Every dependency is license-audited: zero GPL/AGPL in the tree.
Open source is only a promise if you can check it. Every package is tracked on a public compliance dashboard — signed builds, SBOMs, and a zero-GPL/AGPL dependency tree, each claim linked to evidence. Security & supply chain →
Where the data lives
A hosted product sends your privileged work product across a vendor boundary on every prompt. KAOS does not have a vendor boundary. The diagram below is the architecture, not a marketing claim.
From a document to a checkable answer
A reviewer can follow the chain in either direction. From the cited paragraph back to the AI’s claim, or from the AI’s claim back to the page and bounding box on the source document.
Where to go next
You write Python. You have used DSPy, LangChain, Pydantic-AI, Mirascope, or Instructor. Here is what overlaps with what you already know, and where KAOS does something different.
Run research, drafting, and deal-room review on your own documents, inside your own network, with a citation trail your partners and clients can check line by line.
Built at 273 Ventures.