DevCut

The x402-metered video desk for hackathons.

Organizers commission a Challenge Cut (visual spec of what winning looks like). Builders run Submit Ready (HyperFrames / repo / product URL → Devpost-ready MP4). Agents pay per job via x402 — no Runway key paste as the default path.

North star: docs/devcut-thesis.md

Formerly “Director’s Canvas” — same agent pipeline (storyboard → Runway stills → clips → stitch), hackathon-shaped product.

What's here

Route	What it is
/	DevCut landing — three doors
/director	Live storyboard canvas (Challenge Cut / Submit Ready)
/leads	Legacy CopilotKit lead-triage starter
MCP	Expose the agent to Claude / ChatGPT

Quickstart

npx @copilotkit/cli@latest init   # Intelligence
cp .env.example .env
# Set NVIDIA_API_KEY (primary). Optional: VENICE_API_KEY, GEMINI_API_KEY.
# Optionally RUNWAY_API_KEY (else MOCK). GENBLAZE_ENABLED=1 + B2_* for durable exports.
npm install && npm run dev

Open http://localhost:3000 → pick a door (or Run HyperFrames demo) → /director. After stitch: Watch · HyperFrames · Share — download the HF kit.zip for hyperframes init.

Stack

Layer	Technology
Agent	LangGraph + CopilotKit / AG-UI
Planner	NVIDIA → Venice → Gemini (docs/providers.md)
Image / video / audio	Runway (gen4_image_turbo, gen4.5, TTS/SFX)
Composition handoff	HyperFrames (external OS — we feed assets / BRIEF seeds)
Payments	x402 job meter — see docs/x402.md
Durable export	Backblaze B2 via Genblaze when enabled
UI	Next.js + CopilotKit / AG-UI

Docs

DevCut thesis	North star
Demo script	Partner walkthrough
Golden Challenge Cut	Genblaze+B2 brief
Providers	Inference + AG-UI foundation
DevCut × HyperFrames	How we feed HF (BRIEF + assets)
x402 jobs	Pay-per-job API for agents
Hackathon / B2 notes	Genblaze + B2 submission notes
Genblaze feedback issue	Draft issue to file on Genblaze repo
Architecture	Pipeline internals
Roadmap	Shipped / next
Setup	Full local setup

WebMCP integration

DevCut's /director canvas is a genuinely mutable surface — agent actions visibly change what the human sees in real time (shot plan → stills → clips → stitch). We expose that canvas to external agents (e.g. ChatGPT's in-app browser) via the draft WebMCP document.modelContext API, so an agent and a human can work the same cut together.

Design (full rationale + day-by-day plan: docs/webmcp-playbook.md):

1. Start-don't-block. Mutating tools kick off a generation run and return immediately — video pipelines take minutes, WebMCP tool calls must not block. Agents poll get_storyboard_state.
2. Auth-gate mutation. Read-only tools register always (anonymous browsing preserved). Mutating tools register only when signed in (ADR-0002), or always when auth is disabled.
3. One tool = one thing a human already does on the canvas. No kitchen-sink tools; descriptions are written for how agents actually behave.

Tools

Tool	Mode	What it does
get_storyboard_state	read	Brief + every shot (status / error / prompt / urls), whether a run is active, final cut URL. Poll after starting work.
get_export	read	MP4 URLs (final + durable B2), sha256, HyperFrames builder-kit availability.
start_cutdown	mutate	Commission a Challenge Cut from a text brief; starts generation, returns immediately. Fails if a run is already active.
regenerate_shot	mutate	Re-generate one shot by id (after an error or to improve it).
cancel_run	mutate	Cancel the in-flight run; completed shots stay visible.

How it's wired (no BFF / agent changes)

A thin controller singleton decouples the React canvas from browser-context tools — tools cannot reach into component closures:

• apps/frontend/src/lib/webmcp/types.d.ts — ambient typing for the draft document.modelContext API (adjust to the observed runtime surface).
• apps/frontend/src/lib/webmcp/controller.ts — directorController singleton; the page publishes its handlers + latest storyboard snapshot, tools read them back at execute time.
• apps/frontend/src/lib/webmcp/register-tools.ts — the 5 tool definitions (2 read-only + 3 auth-gated mutating), start-don't-block semantics baked into the descriptions.
• apps/frontend/src/app/director/page.tsx — wires the controller + registers tools on mount (DirectorCanvas). The page keeps using its own callbacks, so there is zero human-UX regression.

WebMCP executes in-page with the user's cookies, so BYOK vault decryption, budgetKey(userId, threadId) budgets, and ui_thread_id billing all work unchanged. The isRunning guard is ref-backed (not state captured at effect-registration time) so a stale closure can never report a finished run as "still running" — see "bug class #1" in the playbook.

Try it

1. Enable the WebMCP flag in Chrome (chrome://flags/#enable-webmcp-testing) or use an agent browser that surfaces document.modelContext.
2. Open https://devcut.thisyearnofear.com/director, sign in with GitHub.
3. In the agent, ask it to commission a cut, inspect the storyboard, diagnose a failing shot, regenerate it, then export — all while you watch the canvas.