Hermes agent for lawyers earns a 6.5 out of 10 from me: excellent at keeping work private, weak on legal-specific integrations and outbound safety.

That score is my editorial judgement from the Hermes docs and plugin catalog, not a benchmark I ran.

And to be completely clear, this review is not legal advice; a qualified lawyer reviews everything the agent touches.

Want help deciding whether it fits your practice? Book a free SEO strategy session with my team.

The Scorecard

CriterionScore /10Why, from the docs
Keeping client data local9Local Models: no account, no API key and no network access once a model is downloaded
Sandboxing8Hardened Docker anywhere; the official bubblewrap plugin on Linux only
Access control8Gateways deny unknown users by default, per the security page
Credential handling8The vault encrypts logins locally and keeps them out of the model's view
Firm procedures as skills8/learn turns written procedures into reusable skills
Reading documents6The PDF skill reads the text layer; scans need a separate OCR path
Knowledge base and memory6Knowledge-base skills exist, but built-in memory is tiny and the Memory Wiki is read-only
Deadline reminders6Cron schedules reminders well, but only from dates a person entered
Outbound message safety4Approvals cover risky commands, not emails or chats
Legal-specific integrations3Only de-law and patsnap; no practice-management, signing or e-discovery plugins
Documented certifications2The docs claim no SOC 2, HIPAA or ISO certification

Average those and you land near 6.2; I round up to 6.5 because the privacy rows matter most to a firm.

What Hermes Actually Is For A Law Firm

Hermes is an open-source agent you install on your own computer or server.

It reads files, runs tools, follows saved skills and can work over chat or email gateways.

For a practice, that makes it a private assistant for preparation work: summaries, checklists, research pulls and reminders.

It is not a practice-management system, a document-signing tool or a court-filing service.

I scored it against the jobs a small practice would realistically hand it, using only what the official pages document, and gave each row a one-line reason.

Strength: Data Can Stay On Your Hardware

The FAQ says requests go only to the model provider you configure and that Hermes collects no telemetry.

Conversations, memory and skills live in a local folder.

Pair that with a local model, or a self-hosted endpoint from the providers list, and client text never needs to leave the office.

Few hosted assistants can say that, and it's why the top row scores 9.

It loses a point because the safe choice depends on the firm actually picking a local model.

Strength: Real Guardrails You Can Tighten

The work-machine guide gives a firm a lot of levers.

  • Manual approval mode, so a person sees every flagged command.
  • User-written deny rules that hold even when someone switches approvals off.
  • A write root that blocks file writes outside the folders you list.
  • Secret redaction, already on by default.
  • Hermes secret stores are read-denied to the agent itself.

The guide is honest that these are guardrails, and that containment needs an isolated backend such as Docker or bubblewrap.

I respect a vendor doc that says what its controls don't cover.

On sandboxing specifically, the two routes suit different offices.

Docker containers run with all capabilities dropped, no-new-privileges, a process limit and size-limited temporary storage, and they work on any desktop OS.

bubblewrap is lighter, needs no image or daemon, and hides credentials from each command, but its README limits it to Linux and says it is not a container.

Either one beats running commands straight on a partner's laptop, which is why sandboxing scores 8.

๐Ÿ”ฅ Get the setup trainings behind this verdict Inside the AI Profit Boardroom you get step-by-step video trainings, the installable Agent OS, four coaching calls a week and 3,400+ members testing agents on real business work. โ†’ Join the AI Profit Boardroom

Strength: Gateways And Logins Are Locked By Default

A firm will want to message the agent from a phone, and this is where many tools get loose.

Hermes starts from deny: if no allowlist is set and no allow-all flag is on, the security page says every user is refused.

You then name the people allowed in, per platform or globally, or approve them with a pairing code.

The email gateway has its own sender allowlist, so only the firm's addresses can hand it work.

Logins for research sites or portals go into the credential vault.

The vault prompts once in a masked field, encrypts the password on the machine and fills it into the page without the model seeing it.

That combination earns both the access-control and credential rows an 8.

They lose points only because a careless allow-all flag would undo the protection, and the docs leave that switch available.

Strength: Procedures Become Repeatable Skills

Every firm has written procedures: intake steps, file-opening checklists, letter formats.

Hermes can learn those from a document or a described process and save them as a skill.

For a big source, the skills docs describe a knowledge-base skill with an index and one distilled file per topic, loaded only when needed.

That's the honest version of the "LLM wiki" idea from the video: a knowledge base you build from skills and memory, not a separate wiki feature.

Limit: Document Reading Has Edges

The PDF skill extracts text, tables, metadata and form values from digital PDFs.

It tells you outright that scanned or image-only files belong to its OCR path instead.

The tools reference notes that file reads over about 100K characters are cut and resumed from an offset.

Both are workable, but a lawyer must know them, because a missed page in a long bundle is a missed fact.

Score: 6.

Limit: Memory Is Small And The Wiki Only Reads

Built-in memory holds 2,200 characters of notes and 1,375 characters of user profile.

The official Memory Wiki plugin turns local session history into browsable subject pages and daily logs.

It also gives a read-only audit panel for both memory files, and the README says it writes nothing.

Useful for reviewing what the agent has seen; not a document store for matter files.

Score: 6.

Limit: No Outbound Approval, No Legal Integrations

This is where the score drops hardest.

Hermes approvals cover dangerous commands.

An MCP server marked trust: untrusted in the MCP config asks before write calls, but plain emails and chat replies have no built-in approval gate.

So every client-facing job needs "draft only" written into it.

The plugin catalog has no Clio, docketing, DocuSign or e-discovery entry.

The legal entries are de-law, a community plugin for German and EU law texts, and patsnap, one of Nous's labs plugins at community tier, which its entry says is not supported.

Limit: History Stays Until Someone Deletes It

Every conversation is stored with full history in a local SQLite database, according to the sessions docs.

That is good for privacy, because it stays on your machine, and a duty for the firm, because it stays.

The docs offer hermes sessions prune to delete old ended sessions.

They also warn that compressing a long chat only shrinks the active context and is not a privacy delete.

I found no automatic retention schedule in the docs, so the firm must set its own and run it.

This doesn't get its own row in the table, but it pulls the knowledge-base score down with it.

How It Handles Three Real Small-Firm Jobs

A small litigation firm's chronology

  1. The associate puts text-based pleadings and letters into a matter folder.
  2. Hermes reads each file in chunks and drafts a dated chronology with a source reference per entry.
  3. Anything scanned gets flagged for the OCR path rather than skipped quietly.
  4. The associate checks every entry against the documents before the partner sees it.

A conveyancing practice's completion checklist

  1. A conveyancer saves the firm's completion procedure as a skill.
  2. For each new file, Hermes produces the checklist and marks items it found evidence for in the folder.
  3. The conveyancer ticks items off herself; the agent's marks are only pointers.

A family law solo's weekly reminders

  1. The lawyer keeps a checked deadlines file per matter.
  2. A cron job sends her a due-soon list each Monday into a local file.
  3. She verifies every date against the order before relying on it.

Who Should Use It

  • Solo lawyers and small firms with someone technical enough to own the setup.
  • Practices that want client text on their own hardware with a local model.
  • Firms with written procedures they'd like followed the same way every time.
  • German or EU law practices that want statute text pulled locally through de-law.
  • IP boutiques comfortable sending search queries to Patsnap under their own key.

Who Shouldn't

  • Firms that want a finished legal product with filing, signing and billing built in.
  • Anyone hoping the agent will send client messages without a person reading them.
  • Practices that need a vendor's certification report before adopting a tool; the Hermes docs don't offer one.
  • Teams that would run it on a cloud model by default without a partnership decision.

For a broader score of the same agent across other businesses, read my Hermes Agent for business verdict.

๐Ÿ“ž Want a verdict on your own firm's workflows? Book a free SEO strategy session and we'll go through where an agent fits and where it doesn't. โ†’ Book the free session

Final Verdict: 6.5 Out Of 10

Hermes is the most private way I know to give a small practice an AI assistant, because the model can run on the firm's own machine.

It falls short as a legal product: no practice-management links, no outbound approval and no certifications on paper.

Use it for preparation work a lawyer finishes, keep every message as a draft, and verify every date.

If your firm has the technical owner and the patience to test on dummy matters, the privacy upside is real.

If it doesn't, a 6.5 tool with gaps will demand more care than it's worth to you right now.

The AI Profit Boardroom has the walkthroughs if you decide to set it up.

Also On Our Network

FAQ: Hermes Agent For Lawyers Verdict

Why only 6.5 if the privacy story is so strong?

Privacy rows score 8 to 9, but legal integrations, outbound safety and documented certifications score 2 to 4. The overall reflects both.

Did you benchmark Hermes on legal tasks?

No. The ratings are editorial judgements from the official Hermes docs and plugin catalog, with the reason for each row in the table.

What would raise the score?

A documented approval step for outgoing messages and real practice-management or signing integrations would lift the two weakest rows.

Is the Memory Wiki a document management system?

No. It is a read-only dashboard view of local session history and the two memory files, and it writes nothing.

Can I rely on its deadline reminders?

Treat them as a backup. Cron repeats dates a person entered, so verify every date against the order or rule yourself.

๐Ÿ“บ Video notes + links to the tools ๐Ÿ‘‰

๐ŸŽฅ Learn how I make these videos ๐Ÿ‘‰

๐Ÿ†“ Get a FREE AI Course + Community + 1,000 AI Agents ๐Ÿ‘‰

About Julian

I'm Julian Goldie, an SEO entrepreneur, author and founder of the AI Profit Boardroom, which has 3,400+ members.

I help business owners scale with AI agents, automation and SEO.

  • I've built a 7-figure SEO and link building agency, Goldie Agency, with a team of 70+ SEO experts.
  • I've grown a YouTube channel to 400,000+ subscribers.
  • I wrote "Link Building Mastery".
  • My Udemy courses have taught over 29,000 students.

โ†’ Get my best AI training inside the AI Profit Boardroom

My score stands until the docs add outbound approval or legal integrations, so book a free SEO strategy session if you want help weighing a Hermes agent for lawyers.