What is OpenClaw?
OpenClaw is a free, open-source AI agent framework: software that runs on your own hardware, works around the clock, keeps persistent memory, and executes tasks - shell commands, browser sessions, API calls - in sandboxed processes. It started in November 2025 as Clawdbot, a side project by developer Peter Steinberger, took the name OpenClaw in January 2026, and became the fastest-growing open-source project in GitHub history, passing 347,000 stars as of April 2026.
Two design choices explain the appeal. It is model-agnostic, so it runs on OpenAI, Anthropic, or open-source models - your choice, your keys. And it meets you where you already are: built-in channels include Telegram, Discord, WhatsApp, Slack, and iMessage. You message your agent the way you message a colleague.
Why are lawyers curious about it?
Because OpenClaw is the clearest public demonstration yet that an AI agent can carry real work rather than chat about it. It reads messages, runs research in a browser, calls APIs, and follows up - persistently, around the clock, on hardware you control.
We see the demand directly. When we opened the Referent beta, 450+ applications came in under four weeks for 20 seats. Lawyers are not asking whether agents are coming to legal practice. They are asking which ones to trust.
OpenClaw at a glance
- Price: free and open source
- Origin: started November 2025 as Clawdbot by developer Peter Steinberger; renamed OpenClaw in January 2026
- Traction: 347,000+ GitHub stars as of April 2026 - the fastest-growing open-source project in GitHub history
- Runs: locally on your own hardware, 24/7, with persistent memory
- Executes: “skills” - shell commands, browser actions, APIs - in sandboxed processes
- Models: works with OpenAI, Anthropic, and open-source models
- Channels: Telegram, Discord, WhatsApp, Slack, iMessage
- MCP: can act as a Model Context Protocol client via configuration
- Out of the box: no approval workflow, no audit trail, no legal-specific guardrails
Can lawyers use OpenClaw for client work?
Not out of the box. OpenClaw is a general-purpose personal agent, and a law practice has obligations a personal agent was never designed to meet. Before it touches client work, a firm must add five things.
Confidentiality. Self-hosting keeps data on your machine, which is a real advantage - and it also means the operator owns security patching, access control, and the security of every connected channel. Your duty of confidentiality applies to any tool that touches client data, and with OpenClaw there is no vendor to share that burden.
Supervision. You remain responsible for work product an agent produces, the same way you are responsible for a paralegal’s draft. An agent that acts autonomously at 2 a.m. needs defined review points, or supervision becomes a fiction.
An approval workflow. Out of the box, nothing stands between the agent drafting an email to a client and sending it. Legal work needs a gate: AI prepares, the lawyer signs.
An audit trail. When a client, a court, or your malpractice carrier asks what the agent did and who approved it, “the logs are on a machine in my office somewhere” is not an answer.
Matter context. OpenClaw has memory, but it does not have a system of record. It does not know your clients, matters, deadlines, or billing state unless you connect it to something that does.
None of this is a criticism of OpenClaw’s quality. The engineering is genuinely impressive. The gap is legal operations fit - and the same gap applies to Hermes Agent and every other general-purpose agent framework.
What are the two sane ways to run it in a law practice?
Keep OpenClaw and connect it to a legal system of record over MCP, or use agents built for law practice from the start. The difference is where the guardrails live.
Path 1: connect OpenClaw to Referent over MCP. Referent runs an MCP server - MCP is the open standard that connects AI assistants to tools - at https://mcp.referent.law/mcp. Add that URL to OpenClaw’s MCP client configuration file and complete the sign-in with your Referent account (OAuth; unauthenticated requests get a 401). From then on, OpenClaw can read and act on your firm’s clients, matters, and documents, and every action passes through the same permissions, the same lawyer approvals, and the same audit trail as in the Referent app. OpenClaw stays the front end you like; Referent supplies the guardrails a law practice requires. Setup details for OpenClaw and other MCP clients are at /mcp/.
Path 2: use agents built for law practice directly. Referent’s agents carry intake from first inquiry to proposal and payment link, route email to the right matter and draft replies for approval, track hearings and deadlines against your calendar, and assemble time and activity into billing-ready form - with the approval gate and audit trail built in rather than bolted on. Voice control and iPhone and Android apps included. The security posture is documented at /security/.
Which path fits? If you like running your own infrastructure and already live in OpenClaw, Path 1 gives you legal guardrails without giving up your setup. If you want the outcome without the operations burden, Path 2 is the shorter road. Both lead to the same place: a practice where AI executes the work and the lawyer signs.
Cohort 1 of the Referent beta filled with 450+ applications for 20 seats, and onboarding starts August 10, 2026. The Cohort 2 waitlist is open, and every applicant gets founding-firm perks during the open beta. Join the waitlist.
Frequently asked questions
Is OpenClaw really free?
Yes. OpenClaw is open source and costs nothing to download and run. The full cost of ownership is different - you provide the hardware, you pay for whichever model you connect unless you run an open-source model locally, and you spend the time to configure, patch, and supervise it. For a law firm, operator time is the cost to watch.
Is OpenClaw safe for confidential client data?
It can be, but the responsibility sits entirely with you. Local-first operation means client data stays on hardware you control, which many lawyers prefer to a cloud vendor. It also means you own security patching, access control, and the security of every channel the agent uses. Confidentiality obligations apply to any tool that touches client data, and with self-hosted software there is no vendor sharing that duty.
Does OpenClaw work with Referent?
Yes. Referent runs an MCP server at https://mcp.referent.law/mcp, and OpenClaw can act as an MCP client through its configuration file. After you sign in with your Referent account, OpenClaw can read and act on your firm's clients, matters, and documents. Every action goes through the same permissions, the same lawyer approvals, and the same audit trail as in the Referent app.
Can OpenClaw replace practice management software?
No. OpenClaw has persistent memory, but it is not a system of record - it has no native concept of clients, matters, deadlines, conflicts, or billing. A law practice needs those structures, plus an approval workflow and an audit trail, before an agent's output can be trusted with client work. The workable pattern is an agent connected to a legal system of record, not an agent instead of one.
How is Hermes Agent different from OpenClaw?
Hermes Agent is an open-source autonomous agent from Nous Research, released in February 2026, with an official desktop app since June 2, 2026. Its signature feature is self-improvement - after completing a complex task it writes a structured skill document and reuses it later. Like OpenClaw, it is local-first, and like OpenClaw it ships with no legal guardrails, no lawyer-approval loop, and operator-owned confidentiality. The same two paths apply to it.