What is Hermes Agent?
Hermes Agent is an open-source autonomous AI agent released by Nous Research in February 2026. Its defining feature is self-improvement: after it completes a complex task, it writes a structured skill document, then reuses that skill the next time similar work appears. Add persistent memory across sessions, and the result is an agent that gets better at your recurring work instead of starting from zero every morning.
It is also unusually approachable for an open-source agent. Since June 2, 2026 there is an official desktop app in public preview (v0.15.2) for macOS, Windows, and Linux, so you are not required to live in a terminal. Its gateway reaches CLI, Telegram, Discord, Slack, WhatsApp, Signal, Matrix, email, SMS and more. Everything runs local-first, with no telemetry.
Hermes Agent at a glance
- Released February 2026 by Nous Research; open source
- Self-improving: writes a structured skill document after completing a complex task, then reuses it
- Persistent memory across sessions
- Official desktop app shipped June 2, 2026 (public preview v0.15.2; macOS, Windows, Linux)
- Gateway channels: CLI, Telegram, Discord, Slack, WhatsApp, Signal, Matrix, email, SMS and more
- Local-first; no telemetry
- MCP-compatible: connects to external tools through the Model Context Protocol
- Out of the box: no legal guardrails, no approval workflow, no audit trail
Why does Hermes Agent appeal to a small firm?
Because the skill loop matches how small-firm work actually repeats. A solo practice runs on patterns - the same intake questions, the same engagement letter, the same discovery checklist - and an agent that writes down what it learned on one matter and applies it to the next is compounding in exactly the place where a small firm bleeds time.
The practical details help too. The desktop app removes the command-line barrier that keeps most lawyers away from agent frameworks. Email and SMS in the gateway matter more than Discord, because those are the channels clients actually use. And local-first with no telemetry reads like a confidentiality answer, since your data is not routed through a vendor by default.
The appeal is real and worth taking seriously. It is also incomplete.
What does legal practice require that Hermes Agent does not provide?
Five things: confidentiality controls, supervision, approvals, an audit trail, and matter context. Hermes ships with none of them, by design - it is a general-purpose agent, not legal software.
Confidentiality is broader than no telemetry. Your duty covers the models the agent calls, the channels it reads, and the machine it runs on. Self-hosting means you own security patching and access control, permanently.
Supervision is where the self-improvement feature cuts both ways. Lawyers stay responsible for the work product a tool produces. A skill Hermes wrote for itself after one matter becomes its behavior on the next - and unless you review those skill documents, nobody has checked the reasoning in between. That is not a software flaw; it is an unsupervised junior associate, working at machine speed.
Approvals and the audit trail are the operational versions of the same duty. Nothing in Hermes pauses before a client-facing action to ask you first, and nothing records what happened in a form you could hand to an auditor or a malpractice carrier. Matter context is the last gap: an agent that does not know which client, matter, and deadline a document belongs to cannot file it, bill it, or calendar it correctly.
None of this makes Hermes unusable for lawyers. It defines what has to sit between the agent and your clients.
How do you connect Hermes Agent to your firm’s practice data?
Through the Referent MCP server. Referent runs a server for the Model Context Protocol - the open standard connecting AI assistants to tools - at https://mcp.referent.law/mcp. Add that URL to Hermes’s MCP client configuration file and complete the sign-in with your Referent account. The connection uses OAuth, and unauthenticated requests get a 401.
Once connected, Hermes can read and act on your firm’s clients, matters, and documents - with the same permissions, the same approvals, and the same audit trail as the Referent app itself. A drafted client email still routes to you for sign-off. Every action lands in the log. The agent gains matter context; you keep the controls. The same server works for Claude, ChatGPT, Gemini, Perplexity, and OpenClaw - setup for each client is on the MCP page, and the security model is documented at /security/.
If you are weighing Hermes against OpenClaw for this role, the connection path and the legal caveats are identical for both; the differences between the two agents are covered in the FAQ below.
Should you run Hermes Agent, or use Referent’s agents directly?
Run Hermes if you want a general assistant and you are willing to operate it. That means maintaining the install, reviewing the skills it writes, and supervising its output - reasonable work for a technical solo who enjoys it, a second job for everyone else. With the MCP connection in place, it is a legitimate way to practice.
Use Referent’s agents directly if you want the outcome without the operations. The intake agent takes an inquiry from first contact through qualification, opens the matter, and carries it down to proposal and payment link. The email agent routes correspondence to the right matter and drafts replies for your approval. The deadline agent tracks hearings and deadlines and syncs both ways with your calendar, and the billing-prep agent assembles time and activity into billing-ready form. The trust model does not vary: AI prepares, the lawyer approves every client-facing or high-risk action, and every step is logged. Legacy software records the work - Referent executes it, and the lawyer signs.
Referent will not claim to be the best tool for every firm. It is the AI-native pick: the agent layer and the practice data live in one supervised system, a model described at what AI-native means.
Cohort 1 of the Referent beta filled in under four weeks - 450+ applications for 20 seats - and onboarding starts August 10, 2026. The Cohort 2 waitlist is open at /apply/, and every applicant gets founding-firm perks during the open beta. If Hermes brought you here, connect it. If you would rather skip the setup, apply.
Frequently asked questions
Is Hermes Agent free to use?
The software is open source, released by Nous Research in February 2026, so there is no license fee for the agent itself. You supply everything around it - the hardware it runs on, access to the AI models it calls, and the time to configure and maintain it. For a law firm, the real cost is operational: patching, monitoring, and supervising what the agent does with client data. Budget for that before comparing it to paid tools.
Is Hermes Agent safe for confidential client work?
Local-first operation and no telemetry are real advantages, because by default your data is not flowing to a vendor's analytics. But confidentiality duties cover the whole setup - the models the agent calls, the channels it reads, and the machine it runs on. Hermes ships with no legal-specific guardrails, so nothing stops it from acting on privileged material unless you build that boundary yourself. Connecting it to practice data through the Referent MCP server puts those actions behind your firm's permissions, approvals, and audit trail.
How do I connect Hermes Agent to Referent?
Add the Referent MCP server URL, https://mcp.referent.law/mcp, to the agent's MCP client configuration file, then complete the sign-in with your Referent account. The connection uses OAuth, and unauthenticated requests get a 401. Once connected, Hermes can read and act on your clients, matters, and documents with the same permissions, approvals, and audit trail as the Referent app. Setup details for every major MCP client are on the /mcp/ page.
How is Hermes Agent different from OpenClaw?
Both are open-source autonomous agents that run local-first with persistent memory, and neither ships legal guardrails, an approval workflow, or an audit trail. OpenClaw, started in November 2025, is the larger project - the fastest-growing open-source project in GitHub history, with 347,000+ stars as of April 2026 - and executes its skills in sandboxed processes. Hermes Agent, released February 2026, stands out for writing its own skill documents after complex tasks and for its official desktop app, shipped June 2, 2026. For legal use the caveats are identical: the operator owns confidentiality, security, and supervision.
Does Referent replace Hermes Agent?
They sit at different layers. Hermes is a general-purpose agent you operate yourself; Referent is practice management for solo and small firms, with agents already built for legal work - intake, email routing, deadlines, billing prep - each running under lawyer approval with a full audit trail. If you want both, the Referent MCP server lets Hermes work with your firm's data under those same controls. If you want neither the hosting nor the configuration, Referent's agents run without any of it.