We build the agents, then we stay to run them.

We work with law firms and other paper-heavy businesses, where the day's work arrives as mail, faxes, scans and forms. Most automation projects end at the demo. We design agents around the way your team already works, connect them to the systems you already use, and then operate, monitor and tune them as a managed service.

The night's mail, routed to a sign-off line A fax at 11:42 PM is split into medical records and a cover letter, and an email at 1:07 AM brings a notice of hearing. The agent proposes a folder for each document, and each proposal waits at the sign-off line until a person approves it. Fax, 14 pages 11:42 PM Email 1:07 AM Sign-off 1874 › Medical records 1874 › Correspondence 1847 › Correspondence 2291 › Court notices 2291 › Pleadings Unmatched mail Medical records Cover letter Notice of hearing
Agent's route and proposalApproved by a person3 waiting for sign-off

The morning queue

This is the part your staff see. The agents did the reading overnight, and each proposal waits for a person. Approve one, or change the folder first.

An example with invented documents. In practice these arrive as one-click approvals in Slack or email.

  • Received by fax 11:42 PM

    Medical records, 12 pages

    File to
    Matter 1874 › Medical records
    Notify
    Case manager
  • Received by fax 11:42 PM

    Cover letter, 2 pages

    Agent's note: the letter gives a client name but no matter number, so this folder is the closest match on the name.

    File to
    Matter 1847 › Correspondence
    Notify
    Case manager
  • Received by email 1:07 AM

    Notice of hearing

    File to
    Matter 2291 › Court notices
    Calendar
    Hearing, Thursday at 9:30 AM
    Notify
    Assigned attorney

The record

  1. Fax received, 14 pages. Split into two documents.

  2. A folder is proposed for each document.

  3. Notice of hearing received by email. Hearing date read and a calendar entry drafted.

Each row keeps the source document, the agent's proposal and the name of the person who approved it.

For a law firm, the agents read incoming mail, faxes and court notices, propose where each document belongs and who needs to see it, and then wait. Staff approve each filing with a click. Nothing is filed or sent without that approval.

The approval is enforced in code

An instruction that tells a model to wait for approval can be ignored, or talked around by a cleverly worded document. So that rule does not live in the agent's instructions. Every filing passes through a gate that Midnira wrote, and the agent has no other way to file. The agent proposes. Code checks that a person approved.

Approval
A filing goes ahead only when a person your firm has named as an approver has signed off. Until then the proposal waits.
Signature
Every approval is signed where your staff give it, and the gate checks that signature before it trusts anything else in the approval. The agent holds no signing key, so it cannot approve its own work.
Exact match
An approval covers the proposal as that person saw it and the document as it arrived. If either changes afterwards, the filing is refused until a person approves it again.
Your correction
When your staff change the folder, their choice is what gets filed. A filing only goes into a folder that already exists.

gate.pyThe check every filing goes through

# ---- a person's approval ----------------------------------------------
if approval is None:
    return Decision(HOLD, pid, "no approval on record")
# only the approval channel can sign one; the agent holds no key
if not signed_by(approval, policy.keys):
    return Decision(DENY, pid,
                    "approval is not signed by the approval channel")
who = approval.approver
if approval.proposal != pid:
    return Decision(DENY, pid, "approval names a different proposal", who)
if who not in policy.approvers:
    return Decision(DENY, pid, f"{who} is not a named approver", who)
# an approval covers the proposal as that person saw it, nothing later
if approval.seen != proposal.digest():
    return Decision(DENY, pid, "proposal changed after it was approved",
                    who)
# ---- the document, as received ----------------------------------------
if doc_sha256 is None:
    return Decision(DENY, pid, "document cannot be read", who)
if doc_sha256 != proposal.sha256:
    return Decision(DENY, pid, "document changed since the agent read it",
                    who)
# ---- once only --------------------------------------------------------
if filed is None:
    return Decision(DENY, pid, "record cannot be verified", who)
if filed:
    return Decision(DENY, pid, "already filed", who)
# the approver's folder wins over the agent's; neither can invent one
folder = approval.folder or proposal.folder
if not has_folder(folder):
    return Decision(DENY, pid, f"no such folder: {folder}", who)
return Decision(ALLOW, pid, f"approved by {who}", who, folder,
                corrected=folder != proposal.folder)

ops-fileWhat the agent gets back

$ ops-file fax-0412-b
[HELD] fax-0412-b :: no approval on record
        document: Cover letter, 2 pages
        proposed: Matter 1847/Correspondence

$ ops-file fax-0412-b
[REFUSED] fax-0412-b :: approval is not signed by the approval channel
        document: Cover letter, 2 pages
        proposed: Matter 1847/Correspondence

$ ops-file fax-0412-b
[FILED] fax-0412-b :: approved by case.manager@example.com
        document: Cover letter, 2 pages
        filed to: Matter 1874/Correspondence (the approver changed the folder)

$ ops-file mail-0107
[REFUSED] mail-0107 :: proposal changed after it was approved
        document: Notice of hearing
        proposed: Matter 2291/Pleadings

The code is copied from the gate as written, and the output is the gate's own, run on the example night above. Before the second run, the agent wrote an approval for itself. Before the fourth, it edited a proposal that had already been approved.

A record that checks itself

One approval files one document, once. Every decision goes into the record, the holds and refusals too.

Chained
Each row carries a fingerprint of the row before it. A row that is changed afterwards no longer fits the rows around it.
Anchored
The latest fingerprint goes to you with each report. A record rewritten later, however carefully, no longer matches the fingerprint you hold.
Fails closed
If the record does not check out, the gate stops filing until a person has looked.

The gate's tests run without touching a client's systems.

record.pyEach row is checked against the row before it

for n, line in enumerate(lines, 1):
    try:
        row = json.loads(line)
        prev, proposal, verdict = (row["prev"], row["proposal"],
                                   row["verdict"])
        if not isinstance(proposal, str):
            raise TypeError(proposal)
    except (ValueError, KeyError, TypeError):
        return stop(f"row {n} cannot be read")
    if prev != head:
        return stop(f"row {n} does not follow the row before it")
    if verdict == FILED:
        filed.add(proposal)
    head = _hash(line)
    heads.append(head)

ops-recordWhat a changed record looks like

$ ops-record verify
record OK: 6 rows, 2 filed
head: c976b44297d43ed60431caba4a05d83895f2410c2ec555ba4c420491bc391c9e

$ ops-record verify
record BROKEN: row 3 does not follow the row before it (2 rows verified before it)

$ ops-record verify
record OK: 5 rows, 2 filed
head: 7a04d4db68bd0e3af0125e598e46b6055236ea42cb1a9374db0b197f8788988d

$ ops-record verify --anchor c976b44297d43ed60431caba4a05d83895f2410c2ec555ba4c420491bc391c9e
record BROKEN: it does not contain the anchored head, so rows were changed or removed after that head was issued

Copied and run the same way. Before the second check, one row was edited by hand. Before the last two, a row was removed and every later fingerprint recomputed: the chain alone cannot see that, and the fingerprint from the first check can.

A sample audit record

This is the record one night leaves behind. Each row names the source, who proposed the action and who decided. The data is invented.

TimeWhat happenedBySource
11:42 PMFax received, 14 pages. Split into two documents.Operations agentFax line, job 0412
11:44 PMProposed: file the medical records to Matter 1874 › Medical records and notify the case manager.Operations agentFax 0412, pages 3 to 14
11:44 PMProposed: file the cover letter to Matter 1847 › Correspondence. Flagged: the letter gives no matter number.Operations agentFax 0412, pages 1 to 2
1:07 AMProposed: file the notice of hearing to Matter 2291 › Court notices and add the hearing to the calendar.Operations agentEmail from the court clerk
8:04 AMCorrected: the cover letter goes to Matter 1874 › Correspondence. The correction is saved as a rule.Case managerApproval in Slack
8:05 AMApproved: the medical records are filed as proposed.Case managerApproval in Slack
8:05 AMApproved: the notice is filed and the hearing is on the calendar.Case managerApproval in Slack

Open the full sample record (CSV)

What the agents take on

Intake and triage
Email, fax, scans and portal notices are read as they arrive and sorted by what they are and who needs them.
Document filing
Each document is named, classified and proposed for the right folder in your system of record.
Deadlines and routing
Dates are pulled from notices and letters, sent to the responsible person and proposed for the calendar.
Approvals where you work
Proposed actions show up as one-click approvals in Slack or email, so staff stay in control without a new app to learn.
Monitoring
Every scheduled job is watched. A missed run raises an alert instead of going unnoticed.
Tuning to your procedures
Corrections from your staff are turned into rules, so the same mistake is not repeated.

How an engagement runs

  1. Map

    About two weeks

    We document one workflow as it runs today, including the written procedures and the unwritten ones.

  2. Shadow

    Two to four weeks

    The agents run beside your staff and propose without acting. We compare their proposals with what your people did until the two agree.

  3. Operate

    Ongoing, month to month

    Approvals go live. We run the agents, watch them, report on them and keep tuning as your procedures change.

What we need from you

  • One workflow to start with, and the person who knows it best
  • Your written procedures for it, and the exceptions people handle from memory
  • Access to where the work arrives (inbox, fax, scanner) and to the system it is filed in
  • A named person on your staff who approves the agents' proposals

How we price

A one-time fee to build
Mapping a workflow and building its agents is a fixed fee, quoted in writing before work starts.
A flat monthly fee to run
Operating, monitoring and tuning the agents is a set amount each month. There are no per-document charges and no long-term contract.
What sets the price
The number of workflows, how much arrives each day and how many systems the agents connect to.

The questions firms are told to ask, answered.

State bar guidance tells law firms what to ask before client information goes to an AI vendor. These are our answers. Where an answer depends on your setup, it is written into your agreement before any work starts. This describes our service and is not legal advice.

Is our data used to train AI models?
No. The agents run on Claude, the AI model made by Anthropic. The documents and messages that pass through the agents are not used for model training.
Who can see our documents?
One person at Midnira: the founder, who builds and runs your agents. There are no subcontractors and no offshore staff. Anthropic processes the text the agents read, as the model provider.
Where does our data live, and who else handles it?
Set in your agreement. It names where the agents run, where the logs are kept, the AI provider plan in use and every other provider that touches your data.
How long is it kept, and what happens when we stop?
Set in your agreement. It states how long documents and logs are kept. When the service ends, your data is returned to you and then deleted, and we confirm the deletion in writing.
Do you claim any rights in what we send you?
No. Your documents and data stay yours.
Will you sign a confidentiality agreement?
Yes. Confidentiality terms are part of every agreement, and we sign a business associate agreement when health records are involved.
Will you tell us about a breach, or a demand for our data?
Yes. We tell you within 72 hours of confirming a security incident that affects your data, and promptly if anyone demands access to it, unless the law forbids that.
What security certifications do you hold?
Midnira holds no SOC 2 report and no ISO 27001 certificate. It is a small practice run by a penetration tester who has earned eight GIAC certifications. What we show you instead is how the agents are built, where they run and the record of everything they do.
What do the agents do, and how do we supervise them?
They read, sort and propose. Nothing is filed, sent or put on a calendar until a person on your staff approves it. Each proposal shows its source document, and every action is logged with the person who approved it.
Do we need to tell our clients?
That is your firm's decision under your bar's rules. We give you a plain description of what the agents do and what data they handle, so you can decide and describe it accurately.
How is it billed?
A one-time fee to build and a flat monthly fee to run, both quoted in writing before work starts.

Tell us what you want handled, or what you want tested.

Send a few lines about the work, or book a 30-minute call. A person reads every message and replies.