One Runtime, Many Agents

Leviathan

The infrastructure layer the AI agent ecosystem doesn't know it needs yet

1.24M
Lines of Code
40
Plugins
65
Agent Platforms
9
Graph Layers
6
SDK Tiers

Agent Sprawl Is Killing You

65 coding agents. Each with their own config format, invocation flags, and mental model. You're not building software anymore — you're managing a zoo.

4

Runner Registries

ntm has one. lev harness has another. Poly has a third. ClawTeam built a fourth. None of them talk to each other. Four sources of truth equals zero sources of truth.

3

ExecOptions Interfaces

core/harness has 30+ fields. core/exec has semaphore + eventLogPath. core/domain has timeout + cwd + env. The SDK-first design has profile + prompt + model. None of them match.

3

Invisible Tiers

The DNA contract says use src/handlers/. The CLI discovery only scans src/commands/. Three entire tiers of your own SDK — Research, Graph, Swarm — are unreachable from the CLI.

"You don't have a tools problem. You have a unification problem."

One Method. Every Agent.

Leviathan is an SDK that produces surfaces — not a CLI that happens to have an API.

// This is the entire execution model const result = await lev.exec({ profile: "sdlc.flowmind.exec", prompt: "Fix the auth middleware", }); // Profile resolves runner, model, context, middleware // Poly projects this onto every surface automatically:
>_

CLI

lev exec --profile=...

</>

MCP

mcp__lev__exec()

↔

gRPC

lev.Exec(request)

{ }

REST

POST /api/exec

⚙

Fractal Config

Inline key, domain file, or domain folder glob. config.yaml > {domain}.yaml > {domain}/*.yaml. Your config scales with your complexity — not against it.

↻

Plan-Act-Verify

Every flow reduces to the same substrate: INGEST → OBSERVE → PROPOSE → GATE → APPLY → UPDATE → EMIT. Universal tick cycle. Universal composability.

★

Profile System

Named bundles: runner + model + context + middleware. Dotted namespaces: sdlc.flowmind.exec. Stop remembering 12 flags. Start naming your intent.

Every Website, Agent-Ready

The WebMCP Shim turns any website into a structured tool API your AI agent can call. No scraping. No screenshots. No browser restart. Community-contributed .webmcp.yaml definitions map semantic actions to DOM operations — like Greasemonkey scripts, but for the AI age.

Read the full WebMCP vision →

Screenshot Scraping WebMCP Shim
Token cost ~2,000 / step ~50 / step 40x
Speed 3–5 seconds <100ms
Cloudflare Blocked Passes
Auth Cookie extraction ceremony Inherits browser session
AI required? Yes (vision model) No (structured tool calls)
Read

WebMCP

WebMCP Shim

Agent reads structured data from any website through semantic tool calls.

callTool("get_messages")
~50 tokens / call
Write

GenUI

Generative UI

Agent generates and injects UI components directly into the page.

renderComponent("chart", data)
~100 tokens / component
Control

CDP

Chrome DevTools Protocol

Agent manipulates the DOM directly through the browser's native debug interface.

click(@e3)
~50 tokens / action

Live Now

Running code, not a pitch deck. Explore the system.

Live →

WebMCP Vision

Browser as a structured tool API for autonomous agents. The full architecture and protocol stack.

/webmcp-agentping/
Live →

Heartbeat Dashboard

Live system status and runtime telemetry at a glance. Real-time health of every subsystem.

/genuievolve/
Live →

Constraint Engineering

The seal that only tightens. Ratchet constraints, validation gates, and graph axioms that make agent chaos impossible.

/constraints/
Coming Soon

AgentPing

Human-in-the-loop interaction protocol for AI agents. Approvals, selections, and async feedback.

/agentping/
Coming Soon

Runtime

Poly surface architecture powering every execution path. CLI, MCP, gRPC, REST from one handler.

/runtime/
Coming Soon

Showcase

Live demos of agents, flows, and runtime capabilities. See the system in action.

/showcase/