"Code has changed. It's an endless series of refactors, managed by AI agents and machines."
Liquid Ops is a Tactical Code Intelligence System. It transforms the abstract task of "software engineering" into a visual, spatial, and strategic Real-Time Strategy (RTS) game.
Instead of managing 20 terminal windows or scrolling through flat file trees, you become the Commander of a specialized AI unit. You look at a 3D tactical map of your codebase, identify high-value targets (technical debt, bugs, missing features), and deploy elite autonomous agents to neutralize them.
We believe that Agentic AI requires a new interface. Chatbots are linear. Codebases are high-dimensional graphs. To manage a swarm of intelligent agents effectively, you need Situational Awareness, not just a log stream.
Existing multi-agent systems solve the problem of parallelism by brute force:
- Spawning 20 copies of the repo.
- Running 20 isolated agents.
- Merging the chaos at the end.
- Interface: A spreadsheet or list of logs.
We solve the problem with Shared Intelligence.
- One Truth: All agents share a live Liquid Shadow intelligence graph.
- Surgical Precision: Agents don't overwrite each other; they operate in assigned Sectors (Logic Layer, Data Layer, etc.).
- Interface: A 3D Tactical Map where you see the architecture, the agents, and the mission status in real-time.
Liquid Ops is composed of three primary pillars:
The core of Liquid Ops is Liquid Shadow, an advanced code intelligence system exposed as a Model Context Protocol (MCP) server.
- Micro-Topography: Maps your code into discrete layers (Entry, Logic, Data, Utility).
- Polyglot Tracing: Traces execution flow across languages and files.
- Benefit: Agents don't search blindly. They know where the Auth Logic is because the map tells them, avoiding costly full-repo scans.
We deploy Specialists routed to the best model and CLI for the job. Liquid Ops orchestrates these tools seamlessly:
| Unit | Role | Default Runtime | Behavior |
|---|---|---|---|
| OVERSEER | Commander & Planner | claude |
Session-resumed commander. Prefers live attach when available, otherwise respawns with session resume and continues mission context. |
| RECON | Intelligence | codex / gemini |
Fire-and-forget. Reads massive context, maps dependencies, and identifies risks without mutating code. |
| ENGINEER | Execution | codex / gemini |
Fire-and-forget. Specialized in surgical refactoring, writing features, and fixing bugs. |
Built with Next.js, React Three Fiber (R3F), and Drei, the interface renders your codebase as a Hexagonal Base (The Motherbase).
- Command Spire: The central hub, representing the active project.
- Modules: Components like Logistics, Habitation, and Power generated procedurally based on code statistics (files & symbols).
- Visual Feedback:
- Spires glow & pulse: Indicate active streaming sessions and live chat connections.
- Color Coding: Each imported project receives a unique accent color and codename (e.g., ALPHA, BRAVO).
- Infiltration (Init)
- Launch Liquid Ops.
- Register your local repositories.
- It performs a
shadow_reconscan to index the repository. - A unique 3D Motherbase is instantly constructed on the Command Deck.
- Mission Planning (Briefing)
- Select a project Spire to enter its sector.
- Talk to the OVERSEER. Issue a Directive: "Upgrade all JWT handlers to use the new library."
- Deployment (Action)
- The OVERSEER plans out the execution steps and uses MCP tools to dispatch a RECON or ENGINEER unit to perform the task.
- Real-time updates map the job's progress directly onto your HUD.
- Node.js 20+
- AI CLI dependencies installed globally (
claude,codex,gemini), depending on your preferred runtimes.
- Clone the repository and install dependencies:
npm install
- Start the Command Center development server:
npm run dev
- (Optional) Start the MCP server independently if needed for integration:
npm run mcp
Open http://localhost:3000 to access the Command Deck.
- Frontend: Next.js 16 (App Router), React 19, Tailwind CSS v4, Framer Motion
- 3D Engine: Three.js, React Three Fiber, React Three Postprocessing
- State Management: Zustand (Shared State for Game, Chat, Project, and Performance)
- AI Orchestration: Custom Node Subprocess Handlers capturing execution streams via CLI
- Local Persistence: Better-SQLite3 for structured Agent execution logging (
~/.liquid-ops/sessions/)
Liquid Ops itself exposes an MCP server (mcp-server/index.mjs) used by the OVERSEER to decouple execution. The following tools are available to the commander AI:
recon: Distribute an intelligence gathering job to the RECON unit.engineer: Distribute an execution job to the ENGINEER unit.job_status: Poll on active worker threads.reset_session: Wipe the active context for an assigned repo.
"This isn't just a tool. It's a next-generation command center."