agentos.guide › blog
Hermes agent for lawyers works best inside your Agent OS as a separate, private desk: its own profile, a local model and a short list of jobs that always end in a draft.
This guide gives you the quick start, five prompts to paste and a weekly rhythm.
None of it is legal advice, and a qualified lawyer signs off every output before it is used.
Prefer to plan it on a call? Book a free SEO strategy session with my team.
legal profile so its memory and settings never mix with your other agents.Each move takes one setting or one command, and the docs linked above show the exact keys.
Do those five on dummy files first.
Only move real matters in once the drafts look right.
The Hermes FAQ says requests go only to the model provider you set, with no telemetry, and that your history, memory and skills are stored under ~/.hermes/.
With a local model, that provider is your own machine.
If you want to send work by email, the email gateway has a sender allowlist, so only the firm's own addresses are heard.
If the firm ever chooses a hosted model instead, prompts travel to that company, so make it a deliberate decision.
Logins for research portals belong in the credential vault, which encrypts them locally and keeps them away from the model.
And there are no SOC 2, HIPAA or ISO claims in the Hermes docs, so don't present the setup as certified.
Which sandbox you use depends on the operating system the desk runs on.
On a Linux server, bubblewrap gives each terminal and code command its own sandbox with no image or daemon to manage.
Its README explains that host files are visible read-only, writes are limited to the working directory, and credential stores stay hidden.
Install bubblewrap 0.9.0 or newer, then run hermes doctor to confirm the probe passes.
If the probe fails, the README says commands are never run outside the sandbox, which is the behaviour you want.
On a Mac or Windows laptop, use Docker instead.
The Docker docs describe hardened containers, and the work-machine guide recommends an empty forward list so no secrets are passed into the container.
Write the procedure down in plain words first, then let the agent build the skill.
/learn our client intake procedure in ~/firm/procedures/intake.md.
The skill must: summarise a completed intake form, list missing
documents, flag anything that needs a conflict check by a lawyer,
and save the result as a file in the matter folder. Never contact
the client. Never give a legal opinion.
The skills docs say /learn works from a file, a URL or a procedure you describe.
🔥 Get the Agent OS this desk plugs into
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 building private agents.
Use the pdf skill on ~/matters/2026-014/lease.pdf. First tell me the
page count and whether the file has a text layer. If any page is
scanned, stop and tell me which pages need the OCR path. Then extract
the text in full, reading in chunks until there is no next_offset,
and list every date and obligation with its clause number.
Why so strict? The PDF skill works on the text layer and routes scanned pages to a separate OCR path.
The tools reference also says read_file cuts reads at about 100K characters and returns an offset to carry on from.
/learn ~/firm/knowledge/employment-notes/ as a knowledge-base skill.
Keep an index of topics and one reference file per topic. Note the
source file for every point. Do not include client names.
This is the documented way to get the "LLM wiki" idea from the video.
The skills docs describe large sources becoming a knowledge-base skill: a lean index plus one distilled file per topic, opened only when a question needs it.
To browse what the agent has seen, add the official Memory Wiki plugin.
It turns local session history into browsable pages, shows both memory files read-only, and writes nothing.
Create a cron job for weekdays at 08:00. Read ~/matters/deadlines.md,
which a lawyer maintains, and save a list of items due in the next
10 days to ~/matters/due-soon.md with the source of each date.
Draft only. Do not message anyone.
Cron takes plain-language schedules and can deliver to a local file, a chat or another platform.
The reminder only repeats what the file says, so verify every date yourself against the order or rule.
To pull candidate dates out of a document first, the bundled document-to-action-items skill in the skills catalogue lists obligations and deadlines with citations, which a lawyer then checks before adding them to the file.
For more scheduling patterns, see the Hermes cron job guide.
Using the matter notes in ~/matters/2026-014/, draft a short status
update for the client in plain English. Save it as draft-update.md.
Draft only: do not email, message or upload it anywhere. List any
point you were unsure about at the top for the lawyer.
Hermes approvals cover risky commands, not outgoing messages.
That's why every client-facing prompt says draft only, and the lawyer sends it.
If you later connect an MCP server, set it to trust: untrusted so each write-capable call asks you first.
Asking the agent to list its doubts at the top gives the lawyer a clear place to start the review.
If the firm insists on a hosted model for some drafting, two community plugins try to mask personal details first.
The privacy-gateway plugin runs on Linux only, swaps names and other details for aliases locally, and blocks requests when its detection engine fails to load.
The hermes-pii plugin redacts outgoing user text, but its catalogue entry calls itself fail-open, meaning an error lets the original text through.
For privileged work, prefer the local model and treat masking as an extra layer, never the main one.
| Need | In the catalogue? |
|---|---|
| German and EU law texts | Yes: de-law, community tier |
| Patent search | Yes: patsnap, one of Nous's labs plugins, community tier, "not supported" |
| Session history view | Yes: Memory Wiki, official, read-only |
| Practice management, docketing, e-signature, e-discovery | No entry found |
Keep those missing systems as they are and pass files across yourself.
📞 Want this desk mapped to your firm?
Book a free SEO strategy session and we'll go through your intake, documents and reminders with you.
/learn on new precedent notes so the knowledge base stays current.hermes sessions prune; compression alone doesn't remove anything, as the sessions docs point out.Members of the AI Profit Boardroom get the Agent OS and walkthroughs for wiring this up.
Memory and settings are scoped per profile, so client-related notes never load into your marketing or personal agent.
You can, but prompts then go to that provider. For client material, use a local model or a self-hosted endpoint.
No. It shows browsable pages built from local session history plus a read-only view of the memory files, and it writes nothing.
The PDF skill reads the text layer and sends scanned or image-only pages to its OCR path, so prompt 2 tells the agent to stop and flag them.
No. It is a software setup guide. A qualified lawyer reviews every output and decides what is used.
Keep the profile separate, the model local and the lawyer in charge, and join the AI Profit Boardroom to build your own Hermes agent for lawyers.
About Julian
I'm Julian Goldie, an AI entrepreneur, SEO expert and the founder of the AI Profit Boardroom, which has 3,400+ members.
I help business owners scale with AI agents, automation and SEO.
My YouTube channel has 400,000+ subscribers, and I run Goldie Agency, a seven-figure SEO agency.
→ Hermes Agent For Lead Generation: 5 Prompts In Your Agent OS
→ Grok Bot For Lead Generation: A Prospect Engine In Your Agent OS
→ Hermes Agent For Content Creation In Your Agent OS
The installable system behind every guide on this site — dashboard, agents, memory and pipelines — updated as new tools land, with 3,400+ members and four coaching calls a week.
Join the AI Profit Boardroom →📺 Video notes + links to the tools 👉
🎥 Learn how I make these videos 👉