Code&CounselOS

Inspiration

I'm a solo legal professional, and I built Code&CounselOS because the tools available to small firms and solo practitioners were never designed for us. The legal tech market caters to big law — expensive platforms with complex onboarding, feature bloat for workflows we don't run, and pricing that assumes a 200-attorney firm is footing the bill. Meanwhile, solo attorneys and small firms are wearing every hat — litigator, researcher, drafter, office manager, billing clerk — and duct-taping together free trials and spreadsheets to keep up. When I saw what Gemini 3 could do with multi-step reasoning and structured tool use, I knew this was the moment to build what small practices actually need: not a watered-down version of big-firm software, but a purpose-built AI operating system that matches how we actually work.

What it does

Code&CounselOS is a single AI-powered workspace where solo attorneys and small firms manage their entire practice. Instead of switching between case management, research databases, document editors, and billing tools, everything lives in one place — powered by a team of 12 specialized AI agents orchestrated by Google Gemini 3.

The attorney defines their firm's Standard Operating Procedures — how intake works, how research gets done, how documents get drafted and reviewed — and the system executes them. A Gemini 3 Pro orchestrator reads each request, selects the right SOP, and delegates across specialist agents running on Gemini 3 Flash: intake and conflicts screening, case law research with live CourtListener citations, contract drafting, docketing, discovery, negotiation tracking, billing, and e-filing. A dedicated Quality Control agent powered by Gemini 3 Pro reviews every output before it reaches the attorney. Mission Control lets you watch agents work in real time — monitor status, pause execution, or redirect mid-workflow. The Template Library stores your firm's master forms (NDAs, motions, engagement letters) so agents always draft from your playbook. Activity tracking timestamps every agent action with token costs and projected billable revenue, making AI work recoverable on your invoices. The Analytics dashboard surfaces ROI, time savings, and operational costs at a glance.

How we built it

I built this with my partner, using Google AI Studio with Gemini as my primary development partner throughout the process, with assist from additional coding models when needed. The frontend is React and TypeScript with Zustand for state management, giving the workspace its adaptive split-view layouts. The backend runs on Supabase for multi-tenant data — matters, documents, agent runs, billing entries — with real-time Postgres subscriptions so the UI updates live as agents execute. The agent orchestration layer implements the WAT framework (Workflows, Agents, Tools): SOPs are defined as markdown documents, the Gemini 3 Pro orchestrator parses them into executable steps, and specialist agents on Gemini 3 Flash carry out the work using a registered tool inventory that includes live legal APIs. The two-tier model strategy — Pro for reasoning-heavy orchestration and QC, Flash for high-throughput specialist execution — was a deliberate architectural decision to keep costs realistic for small practices.

Challenges we ran into

The biggest challenge was making agent orchestration predictable enough for legal work. Attorneys can't tolerate hallucinated citations, missed deadlines, or unauthorized filings. I built structured SOP execution with explicit approval gates for high-stakes actions — the system asks before it acts on anything irreversible like e-filing or fund movements. Getting the Bullpen UI (Mission Control) to surface real-time agent status, cost telemetry, and billable time without overwhelming a solo practitioner took several iterations. Balancing automation with human control was a constant tension: too much autonomy and attorneys lose trust, too little and you've just built a fancy chatbot. I also had to design the parallel execution logic carefully — when two agents can run simultaneously (researching two cases) versus when they need sequential handoffs (research must finish before drafting begins), and how the orchestrator manages context between them.

Accomplishments that we're proud of

The SOP-driven architecture. Attorneys define the rules, agents follow them. This isn't a generic AI assistant — it's a system that respects how legal professionals actually structure their work. Every workflow is a firm-authored procedure, not a black-box prompt.

Making agent work billable. Every agent action is time-stamped, cost-tracked, and mapped to a matter. The Analytics dashboard projects billable revenue from AI work, showing small firms exactly how agent output translates to recoverable value on their invoices. This turns AI from an expense line into a revenue multiplier.

Building it with just two of us in a hackathon timeframe. One person — full-stack, legal domain expertise, and engineering, and the other person UI/UX experience — producing a functional multi-agent legal workspace with live case law integration, real-time orchestration, template management, and billing analytics. That's exactly the point of Code&CounselOS: one or two people, equipped with the right AI architecture, can build what used to require a team.

What we learned

Building a multi-agent system taught me that the hard problem isn't getting AI to generate text — it's getting agents to collaborate reliably under defined procedures. Designing the WAT framework forced me to think like both an engineer and a managing partner: which tasks can run in parallel, which need sequential handoffs, where does a human need to stay in the loop, and how do you surface all of that without burying the user in complexity.

I also learned how much the two-tier Gemini 3 model strategy matters in practice. Early prototypes ran everything on Pro — accurate but expensive. Moving specialist work to Flash while keeping orchestration and QC on Pro cut projected costs significantly without measurable quality loss on routine tasks. For small firms watching every dollar, that architectural decision is the difference between viable and theoretical.

What's next for Code&CounselOS

The next milestone is a fully local version that can run offline using open-source models. Solo attorneys working in courthouses, rural offices, or client sites don't always have reliable connectivity — and some matters require data to never leave the machine. I'm building toward a local-first architecture where the same SOP-driven agent orchestration runs against open-weight models on the attorney's own hardware, with the option to connect to Gemini 3 when cloud capabilities are needed. The goal is to make Code&CounselOS work everywhere a solo attorney works — not just where there's good WiFi.

Built With

• courtlistener
• docling
• gemini
• google
• googleaistudio
• postgresql
• react
• supabase
• superdoc
• typescript
• zustand