The Next Chapter for Kelvin: From Web Platform to Agentic Data Infrastructure

Three Years of Building

When we launched the Kelvin Legal Data OS in 2022, we set out to solve a fundamental problem: legal data lives in disconnected silos, and organizations cannot extract value from information they cannot access, connect, or analyze at scale. Kelvin was our answer: a platform that could ingest, normalize, enrich, and serve legal data from virtually any source.

Over three years, Kelvin processed billions of documents. We built connectors for practice management systems, document management platforms, billing databases, and public legal repositories. We developed legal-specific NLP models, entity extraction pipelines, and knowledge graph infrastructure. We shipped the product to law firms, legal departments, and legal technology companies.

We learned an enormous amount. And what we learned has led us to a significant strategic shift.

What the Market Taught Us

The most important lesson from three years of Kelvin deployments is this: the interface is changing.

When we started building Kelvin, the expected delivery mechanism for data products was a web application: dashboards, search interfaces, visualization tools. Users would log in, run queries, explore results. This is how the legal technology industry has delivered data products for decades.

That model is being displaced. The rise of agentic AI systems means that the primary consumers of structured data are increasingly not humans interacting with a browser, but AI agents interacting with APIs, libraries, and data feeds. When an attorney asks Claude to analyze a set of contracts, or when a developer builds an agent that monitors regulatory changes, the data infrastructure needs to speak the language of agents, not the language of web applications.

We watched this shift happen in real time across our customer base. Organizations that had initially asked for dashboards began asking for API endpoints. Teams that had wanted search interfaces began asking for Python libraries they could import into their agent workflows. The most sophisticated customers wanted to plug Kelvin directly into tools like Claude Code and Cursor through MCP (Model Context Protocol) servers.

The message was clear: the future of legal data infrastructure is not another web platform. It is a data layer for AI agents.

Retiring Kelvin v1

As of Q4 2025, we have retired the Kelvin Legal Data OS v1 web platform. Existing customers have been migrated to API-based access, and the web interface is no longer actively maintained.

This was not a decision we made lightly. Kelvin v1 represented years of engineering effort and served our customers well. But continuing to invest in a web-first architecture would have meant building for the past rather than the future. We owe our customers, and ourselves, a product that meets the moment.

Introducing Kelvin Agentic OS

Kelvin Agentic OS is our second-generation product: a suite of Python libraries, REST APIs, and MCP servers designed to provide AI agents with structured access to legal data and capabilities.

The core philosophy is simple: if an AI agent needs to work with legal data, Kelvin Agentic OS should make that as seamless as importing a library or calling an API endpoint. No web interface required. No human in the loop for data access. Just clean, well-documented, programmatic access to the capabilities that legal AI agents need.

The Module Architecture

Kelvin Agentic OS is organized into 15 focused Python modules spanning the full lifecycle of legal data, from ingestion and extraction through enrichment and delivery. They fall into five categories:

Core & Infrastructure

kelvin-core: The foundation, including a tool registry system, execution engine, prompt templates, configuration profiles, and virtual filesystem. Every other module registers its capabilities through kelvin-core’s @kelvin_tool decorator, creating a unified catalog of 150+ specialized tools that agents can discover and invoke.

Document Processing

kelvin-pdf: Async PDF extraction, rendering, and page splitting built on pypdfium2. Handles the reality that most legal data still arrives as PDF.
kelvin-office: Read, create, and edit Word (.docx) and Excel documents using the Office Open XML standard.
kelvin-markdown: Bidirectional Markdown-to-DOCX conversion with round-trip fidelity, enabling agents to work in Markdown while delivering polished documents.
kelvin-tabular: Polars-based tabular data processing for Excel, CSV, JSON, and database sources, with statistical analysis capabilities.
kelvin-tika-client: Type-safe async client for Apache Tika, providing broad-spectrum document parsing and metadata extraction across dozens of formats.

NLP & LLM

kelvin-nlp: A high-performance NLP toolkit with a Rust core and Python bindings. Handles tokenization, sentence segmentation, POS tagging, and linguistic processing at speeds that matter when you are processing millions of documents.
kelvin-llm: 44+ structured extractors organized into seven complexity layers, from primitives and atomic entities through temporal, quantitative, geopolitical, conditional, and reference extraction. A middleware system provides retry, caching, validation, metrics, and reflection.
kelvin-llm-client: A lightweight, unified client for OpenAI, Anthropic, and compatible APIs with token counting, async support, and CLI tooling.

Data Sources & Research

kelvin-source: Source code and data analysis with Rust-accelerated parsing, deduplication, and optional browser automation.
kelvin-research: Connectors to 15+ legal and regulatory data sources, including SEC EDGAR, GovInfo, USPTO, PACER, and the Caselaw Access Project. Async HTTP clients with consistent schemas.
kelvin-web: 35 web tools including HTTP and browser clients (httpx, Playwright), web scraping, PACER legal search, form extraction, link extraction, and HTML-to-Markdown conversion.
kelvin-billing: LEDES billing data processing, covering invoice parsing, classification workflows, timekeeper extraction, and batch processing pipelines.

Integration

kelvin-mcp-server: Exposes the full Kelvin tool catalog via Model Context Protocol (MCP) over stdio, HTTP, and SSE transports. This is how Claude Code, Cursor, and other MCP-compatible tools access Kelvin natively.
kelvin-solr-client: Type-safe async client for Apache Solr with faceting, filtering, highlighting, and connection pooling.

Each module is a standalone Python package that can be installed via pip, used in any Python environment, and composed with other tools in an agent’s toolkit. Several modules (kelvin-nlp, kelvin-source) use Rust cores with Python bindings for performance-critical operations.

API and MCP Access

For organizations that prefer HTTP-based integration, every module capability is also available through REST APIs with OpenAPI documentation. Rate limiting, authentication, and usage tracking are built in.

For teams building with Claude Code, Cursor, or other MCP-compatible tools, we provide MCP servers that expose Kelvin capabilities as tool definitions. This means an AI agent running in Claude Code can call Kelvin functions as naturally as it calls any other tool, with no custom integration code required.

What Stays the Same

The core capabilities that made Kelvin valuable have not changed:

Legal-specific NLP trained on billions of tokens of legal text
Entity extraction and normalization for legal entities, courts, jurisdictions, and citations
Document processing pipelines for common legal document types
Data enrichment that adds structure and metadata to unstructured legal content

What has changed is how these capabilities are delivered. Instead of a monolithic web application, they are now modular, composable, and designed for programmatic consumption.

What Changes

No web dashboard. Kelvin Agentic OS is a data layer, not a user interface.
Python-first. Every capability is available as a Python library call.
Agent-native. MCP servers and structured API responses are first-class citizens.
Modular. Install only what you need. Compose modules as your use case requires.

Timeline

Q1 2026: Python libraries in beta, available to early access partners
Q2 2026: General availability of all core modules, REST APIs, and MCP servers
Ongoing: New modules and capabilities added based on customer needs and market developments

Get Early Access

We are working with a select group of organizations on early access to Kelvin Agentic OS. If your team is building agentic AI workflows that need structured legal data, we want to hear from you.

Contact us at hello@273ventures.com to discuss early access, or visit 273ventures.com/contact to schedule a conversation.

The era of the web dashboard as the default interface for legal data is ending. The era of agent-native data infrastructure is beginning. We are building for what comes next.