Most legal tech stack guides are logo walls: forty products in a grid, every one of them labeled essential. This is not that. A solo or small firm does not need forty tools. It needs one good answer in each of five categories, connected so they share a single picture of the firm’s work. What follows is the map - the categories, what belongs in each, and the one thing not to buy. It is part of the AI-native firm atlas.
How is an AI-native stack organized?
Five layers. At the center, a system of action that runs the firm’s operations. Beside it, deep-work AI tools for research and drafting. Above it, a personal AI assistant - the one the lawyer already talks to every day. Connecting all of it, MCP, the open standard that lets AI tools act on firm data. Underneath everything, the non-negotiables: security posture, audit trail, approval workflow.
If a tool does not fit one of these layers, it needs a strong argument for existing. Firms of one to ten lawyers - the segment mapped in our breakdown of the American profession by firm size - can typically run on four or five products total.
Layer 1 - what runs the firm?
A system of action: practice management software that executes work rather than recording it. This is the load-bearing choice in the stack, because every other layer reads from and writes to it.
The distinction matters. Legacy practice management is a system of record - a database with a calendar, where the lawyer does the work and then documents that it happened. An AI-native system inverts this. Referent is the worked example, and the AI-native pick for solo and small firms: an intake agent captures an inquiry, qualifies it, opens the matter, and carries it through to proposal and payment link. An email agent routes correspondence to the right matter and drafts replies for approval. A deadline agent tracks hearings and deadlines and syncs both ways with the calendar. A billing-prep agent assembles time and activity into billing-ready form. Voice control lets the lawyer tell Referent what needs to happen. The trust model is fixed throughout: AI prepares, the lawyer approves every client-facing or high-risk action, and every step lands in the audit trail. Legacy software records the work - Referent executes it, and the lawyer signs.
An existing system of record does not have to go. Referent connects to it over API, or migrates the data in - whichever the firm prefers. For the full argument on what AI-native means, see What is an AI-native law firm?
Layer 2 - what handles the deep legal work?
Dedicated research and drafting tools: AI built for long documents and legal reasoning. Contract review, case-law research, first drafts of briefs and memos. Harvey and Legora are the best-known names in this category, and they are genuinely strong at what they do - sustained analytical work on complex legal documents.
The line to draw is scope. These tools produce work product; they do not run the firm. A drafting tool does not chase intake, route email, watch deadlines, or prepare bills. It belongs in the stack next to the system of action, not instead of it. A firm that buys a research tool and calls its stack done has automated the work it bills for and left the operations untouched.
Layer 3 - which personal AI assistant belongs in the stack?
The one the lawyer already uses. Claude, ChatGPT, Gemini, and Perplexity all support MCP, which means any of them can become an interface to the firm’s actual data rather than a general-purpose chat window. The choice between them matters less than connecting whichever one you choose.
For the self-hosting-minded, there is a second branch of this category: open-source agent frameworks that run on your own hardware. OpenClaw - started in November 2025 as Clawdbot, renamed in January 2026, and past 347,000 GitHub stars as of April 2026, the fastest-growing open-source project in GitHub history - runs locally around the clock with persistent memory and is model-agnostic, working with OpenAI, Anthropic, and open-source models. Hermes, from Nous Research, shipped its official desktop app in June 2026 and writes reusable skill documents after completing complex tasks. Both are MCP-compatible. If you are weighing either for a practice, the honest guides cover the legal-operations gap: OpenClaw for lawyers and Hermes Agent for lawyers.
Both also carry the same caveat for legal use: no legal guardrails, no approval workflow, no audit trail out of the box. Self-hosting means the operator owns confidentiality, security patching, and supervision. That is a real cost, and for client work it is the lawyer’s professional responsibility on the line - the duty to supervise work product does not lapse because the supervised party is software.
Layer 4 - what connects the layers?
MCP - the Model Context Protocol, an open standard originated by Anthropic in November 2024 that connects AI assistants to tools and data. It is the reason the stack above is a stack rather than five disconnected subscriptions.
Referent runs an MCP server at https://mcp.referent.law/mcp. Connect it in your AI app’s connector settings - paste the URL, sign in with your Referent account - and that assistant can read and act on the firm’s clients, matters, and documents. Same permissions, same approvals, same audit trail as in the app; unauthenticated requests get a 401. Setup instructions for each client are on the MCP page.
Layer 5 - what are the non-negotiables?
Three things every layer must have before it touches client data: a security posture you can verify, an audit trail, and an approval workflow.
Security posture is checkable. For Referent, the specifics are public on the security page: AES-256 encryption at rest, TLS 1.2+ in transit, per-firm isolated workspaces, role-based access, no training on firm data, SOC 2 and ISO 27001 in preparation, CASA verified. Ask every vendor in your stack for the equivalent page. If it does not exist, that is the answer.
The audit trail and approval workflow follow from professional obligations rather than preference. Lawyers remain responsible for supervising work product, and confidentiality duties extend to every tool that touches client data. An AI layer without an approval step asks the lawyer to sign work unseen; one without an audit trail makes supervision unprovable. Neither is compatible with running a practice.
What should you not buy?
Tools that fragment context. The failure mode of most legal tech stacks is not a missing tool - it is six tools each holding a private copy of the truth: a CRM that does not know about matters, a drafting tool that cannot see the client file, a calendar that is not the deadline system.
The test for any purchase is one question: does it read and write the firm’s shared record, over API or MCP, or does it keep its own? A brilliant tool with isolated context makes the stack worse, because every piece of data it captures is a piece the rest of the stack can no longer see. Buy fewer tools, connected, over more tools, siloed.
Where to start
Pick the system of action first - it is the layer everything else plugs into - then connect the assistant you already use over MCP, and add deep-work tools when the practice demands them. Referent’s Cohort 1 filled in under four weeks, with 450+ applications for 20 seats. The Cohort 2 waitlist is open at /apply/, and every applicant gets founding-firm perks during the open beta.
Frequently asked questions
Do I need to replace my existing practice management software to go AI-native?
No. Referent connects to an existing system of record over API, so the current database can stay where it is while agents execute the day-to-day work. Firms that prefer a clean break can migrate their data into Referent instead. The practical order is to add the system of action first and decide the fate of the legacy system after the workflows have moved.
What is MCP and why does it matter for a law firm?
MCP (Model Context Protocol) is an open standard, originated by Anthropic in November 2024, that connects AI assistants to tools and data. It matters because it turns a general-purpose assistant into one that can read and act on the firm's actual clients, matters, and documents. Referent runs an MCP server at https://mcp.referent.law/mcp, and connecting it requires signing in with a Referent account. Actions taken through MCP carry the same permissions, approvals, and audit trail as actions taken in the app.
Are open-source agents like OpenClaw and Hermes safe for legal work?
They are capable, and they are not supervised by default. Neither ships with legal guardrails, an approval workflow, or an audit trail out of the box, and self-hosting means the operator owns confidentiality, security patching, and oversight. A lawyer's duty to supervise work product applies regardless of the tool. The workable pattern is to connect them to systems that enforce approvals and logging - over MCP, an agent acting on Referent data inherits the same approvals and audit trail as any other client.
Where do Harvey and Legora fit in this stack?
They are deep-work tools - layer two of the map. Both are known for sustained research and drafting work on complex legal documents, and that strength is real. What they do not do is run the firm: intake, email routing, deadlines, and billing preparation sit in the system-of-action layer. The two layers complement each other rather than compete.
How many tools does a small firm actually need?
One good answer per layer: a system of action, a research and drafting tool if the practice area demands one, a personal AI assistant, and MCP connections between them - with a verifiable security posture across all of them. That is typically four or five products. More tools usually means more fragmented context, not more capability. The test for any addition is whether it reads and writes the firm's shared record.