[ TALKING IS THE EASY HALF ]
Every voice vendor can build a phone agent that talks. Almost none can build one that writes, safely, into your practice management system. We already shipped that half: scoped tools into Clio, MyCase, Filevine, and IntakeQ, human confirmation on every write, and an append-only audit log.
30 min with a co-founder. Rough quote in 3 business days.
4
open-source connectors on npm
2,003
npm downloads in the last 30 days
3
HIPAA audits passed on systems we built
[ ANATOMY OF AN AI VOICE AGENT ]
Telephony, speech-to-text, and text-to-speech are components you can buy and swap in an afternoon. What decides whether the agent is usable in a law firm or a clinic is the layer underneath.
01
Phone number, speech-to-text, text-to-speech. We assemble these from mature vendors and keep them swappable, so you are never locked to one provider's voice or pricing.
02
Claude via the API with training disabled and Zero Data Retention, or a self-hosted model where policy requires it. The call script, the questions, the escalation rules. Model-agnostic by design.
03
Scoped tools into Clio, MyCase, Filevine, or IntakeQ. The agent can create the matter and hold the consult slot. It cannot touch anything you did not give it a tool for.
04
Writes that need judgment wait for a human. Every tool call, every input, every approval is logged append-only, so you can answer "what did the agent do at 9pm on Saturday" with a record.
[ WHAT THE AGENT DOES ]
A transcript in your inbox is not intake. The call is only finished when the record exists in the system your team actually works in.
The caller reaches a person-paced voice instead of voicemail. The agent captures who called, what happened, when it happened, and how to reach them. Then it does the work:
The front desk is on one line and three patients are on hold. The agent takes the routine calls and books them into the practice system, not into a spreadsheet someone re-keys later:
[ BOOKING IS THE HARD HALF ]
Demos end at the transcript. Production starts the moment an agent is allowed to change a record in a system that a malpractice claim can be filed against. That is a different engineering problem, and it is the one we have already solved.
You choose which tools exist. A voice agent with a "create matter" tool and no "delete document" tool cannot delete a document, no matter what the caller says to it.
Reads are free. Writes are gated. The agent proposes the matter and the consult slot; a person approves it. Nothing irreversible happens on a phone call at 2am.
Every tool call, input, and approval recorded and immutable. This is what makes an agent defensible in an ethics review or a HIPAA audit, and it is missing from most voice products.
Privilege and PHI decide the design, not the other way around: matter-scoped access, BAA coverage, training disabled, Zero Data Retention, self-hosted where policy demands it.
This is the same architecture we ship for every agent build. See how we build AI agents.
[ READ THE CODE FIRST ]
Four open-source MCP connectors, MIT licensed, published on npm and the official Anthropic MCP Registry under the @oktopeak scope. 2,003 downloads in the last 30 days. They are the tool layer the voice agent calls, public for anyone to inspect before a single sales call.
@oktopeak/clio-mcp. Matters, contacts, documents, tasks, calendar, billing, audit-log export. Listed in the Clio App Directory. Hub page.
@oktopeak/mycase-mcp. The first and only open-source MyCase connector. Hub page.
@oktopeak/filevine-mcp for PI litigation and @oktopeak/intakeq-mcp, the first compliant IntakeQ/PracticeQ connector. Same audit-logged architecture.
[ WHERE WE DEPLOY ]
Voice is general. Privilege and PHI are not. We build where a missed call costs a case or an appointment slot, and where the record has to hold up under audit.
On Clio, MyCase, and Filevine. Matter-scoped access, conflict checks flagged for a human, attorney sign-off gates, audit logging.
Legal AI integration →On IntakeQ, PracticeQ, and your EHR. BAA-covered, Zero Data Retention, clinical questions escalated to a human.
Healthcare AI integration →[ FAQ ]
Voice intake on a system you already run (Clio, MyCase, Filevine, IntakeQ) starts at $1,700. A custom build with its own write-path, confirmation gates, and audit logging runs $17,000 to $53,000 fixed price, depending on how many call flows you need and how many systems the agent writes into.
Discovery is free: a call, a draft scope, and a rough quote before you commit. A standalone audit of an existing intake setup is $2,420, credited toward the build if you go ahead. Retainers start at $2,420/month.
It answers the phone, captures the caller (name, contact details, what happened, when), and then does the part most vendors stop short of: it writes the result into your system. For a law firm that means a matter shell with the parties on it, a conflict check flagged for a human, and a consult booked on the right calendar. For a clinic it means a patient appointment booked or rescheduled in the practice system.
Every write is recorded in an append-only audit log: what the agent did, what data it used, and who approved it.
The speech layer is newer for us than the integration layer, and we will say that plainly on the call. What we have shipped and open-sourced is the hard half: the write-path into Clio, MyCase, Filevine, and IntakeQ, with scoped tools, human confirmation gates, and append-only audit logs. That is the part most voice vendors cannot do.
Telephony, speech-to-text, and text-to-speech are commodity components we assemble and can swap. The code that decides what the agent is allowed to write into your system of record, and proves what it wrote, is ours and is public on npm.
Whichever fits your confidentiality requirements. Most builds run on Claude via the API with training disabled and Zero Data Retention. For organizations whose policies rule out third-party AI processors, we deploy self-hosted models on your own infrastructure.
The architecture (scoped tools, gates, audit log) is model-agnostic. The model is a swappable component, and so is the speech layer.
That is the main reason firms buy it. The agent answers at 9pm on a Saturday, takes the caller through the same intake script your staff would, and leaves a complete record waiting on Monday morning: the caller captured, the matter shell created, the conflict check flagged, the consult held on the calendar.
Nothing irreversible happens without a person. Writes that need judgment sit behind a confirmation gate until someone on your team approves them.
Yes, and you can read the code before you talk to us. We maintain four open-source MCP connectors, MIT licensed, on npm and the official Anthropic MCP Registry under the @oktopeak scope: Clio (26 tools), MyCase (18 tools), Filevine, and IntakeQ. The Clio connector is listed in the Clio App Directory, and @oktopeak/intakeq-mcp is the first compliant IntakeQ/PracticeQ MCP connector.
If you run a different system, the write-path is built the same way against its API.
Architecturally, not with a policy document. For clinics: BAA coverage across the vendor chain, Zero Data Retention on the model so call content is not retained or trained on, and human verification before anything touches the chart. We have passed 3 HIPAA audits on systems we built.
For firms: matter-scoped data access, attorney sign-off before anything client-facing goes out, and an append-only log of every action the agent took. See how we build the agent layer.
One voice workflow on a system you already run ships in 1-2 weeks. A custom build, including the write-path, the compliance controls, and staff training, takes 4-8 weeks to production. Discovery before that is free.
[ GET STARTED ]
A call with a co-founder, not a sales rep. Bring the calls you are missing and the system you book into. We will tell you honestly what the agent can write, what has to stay with a human, and what it costs.
